Embedded Files
April 2026
April 2026
April 30: AERObotics: The Art of Catching Objects in the Air
April 30: AERObotics: The Art of Catching Objects in the Air
Speaker: Prof. Debasish Ghose (Indian Institute of Science, Bangalore )
Speaker: Prof. Debasish Ghose (Indian Institute of Science, Bangalore )
Abstract: A branch of robotics, called mobile robotics, looks at problems related to guiding robots to move over a terrain. Robots in the air are aerial robots and often have similar requirements of guiding themselves from one point to another in the air. A particularly interesting problem is when the robot has to intercept a moving object flying in the air. For instance, how can a robot gripper mounted on a drone catch a ball in the air? In this talk we will see how ideas from the literature on intercepting targets work in the domain of aerial robots.
Abstract: A branch of robotics, called mobile robotics, looks at problems related to guiding robots to move over a terrain. Robots in the air are aerial robots and often have similar requirements of guiding themselves from one point to another in the air. A particularly interesting problem is when the robot has to intercept a moving object flying in the air. For instance, how can a robot gripper mounted on a drone catch a ball in the air? In this talk we will see how ideas from the literature on intercepting targets work in the domain of aerial robots.
Speaker Bio: Prof. Debasish Ghose is an honorary professor at the Indian Institute of Science, Bengaluru. Earlier he was a professor at the same institute in both the Department of Aerospace Engineering and the Robert Bosch Centre for Cyber-physical Systems. He works in the area of guidance and control of autonomous systems, robotics, and game theory. He has held various visiting positions such as at the University of California at Los Angeles, the University of Leicester, and the University of Exeter in the UK. He is a fellow of INAE, NASI, INSA, AAAI, and an associate fellow of AIAA.
Youtube Recording: @Robotics Café
May 2026
May 2026
May 7: Leveraging Predictive Uncertainty for Reliable Model-Based Planning and Control
May 7: Leveraging Predictive Uncertainty for Reliable Model-Based Planning and Control
Speaker: Prof. Arun Kumar Singh (University of Tartu, Estonia)
Speaker: Prof. Arun Kumar Singh (University of Tartu, Estonia)
Abstract: Model-based planning and control remains a dominant paradigm for autonomous decision-making in robotics. However, the underlying models—whether representing robot dynamics or complex environmental interactions—are inherently prone to error. In this talk, I will emphasize the critical need to quantify the uncertainty associated with these learned models and demonstrate how such estimates can be everaged to achieve robust planning and control. I will highlight recent advancements, including my group's latest work on learning uncertainty in a task-aware manner. To ground these theoretical insights, I will showcase practical applications in monocular image-based navigation and autonomous driving.
Abstract: Model-based planning and control remains a dominant paradigm for autonomous decision-making in robotics. However, the underlying models—whether representing robot dynamics or complex environmental interactions—are inherently prone to error. In this talk, I will emphasize the critical need to quantify the uncertainty associated with these learned models and demonstrate how such estimates can be everaged to achieve robust planning and control. I will highlight recent advancements, including my group's latest work on learning uncertainty in a task-aware manner. To ground these theoretical insights, I will showcase practical applications in monocular image-based navigation and autonomous driving.
Speaker Bio: Prof. Arun Kumar Singh is a Professor of Motion Planning at Institute of Technology, University of Tartu, Estonia. He is the PI of the AKS Lab. His researchc interests lie in developing algorithmic foundations of robotics, especially motion planning, control and learning4control. His group consistently publishes in top robotics conferences and journals. He is currently leading a 8M Eruopean project XSCAVE which aims to develop AI based planners and controllers for safety-critical operations involving heavy machineries in Forestry, Logistics and Construction.
Youtube Recording: @ Robotics Café
May 14: Passive dynamics is a good basis for robot design and control. Not.
May 14: Passive dynamics is a good basis for robot design and control. Not.
Speaker: Prof. Andy Ruina (Cornell University, USA)
Speaker: Prof. Andy Ruina (Cornell University, USA)
Abstract: Many airplanes can, or nearly can, glide stably without control. So, it seems natural that the first successful powered flight followed from mastery of gliding. Many bicycles can, or nearly can, when in motion, balance themselves. Bicycle design seems to have evolved to have this feature. Also, we can make toys and ‘robots’ that, like a stable glider or coasting bicycle, stably walk without motors or control in a remarkably human-like way. Again, it thus seems to make sense to use `passive-dynamics’ (uncontrolled dynamics) as a core for developing the control of walking robots and to gain understanding of the control of walking people.
Abstract: Many airplanes can, or nearly can, glide stably without control. So, it seems natural that the first successful powered flight followed from mastery of gliding. Many bicycles can, or nearly can, when in motion, balance themselves. Bicycle design seems to have evolved to have this feature. Also, we can make toys and ‘robots’ that, like a stable glider or coasting bicycle, stably walk without motors or control in a remarkably human-like way. Again, it thus seems to make sense to use `passive-dynamics’ (uncontrolled dynamics) as a core for developing the control of walking robots and to gain understanding of the control of walking people.
That's what I used to think. But, so far, this passive approach has not led to robust walking robots. And, when looked at more carefully, the airplane and bicycle motivations are not as suggestive as they seemed. What about human evolution? Unlike what would be indicated by the passive-dynamic paradigm, human beings did not evolve dynamic bodies and then learn to control them. Rather, people had elaborate control systems way back when we were fish and even when we were worms with no skeletons.
That's what I used to think. But, so far, this passive approach has not led to robust walking robots. And, when looked at more carefully, the airplane and bicycle motivations are not as suggestive as they seemed. What about human evolution? Unlike what would be indicated by the passive-dynamic paradigm, human beings did not evolve dynamic bodies and then learn to control them. Rather, people had elaborate control systems way back when we were fish and even when we were worms with no skeletons.
Now, instead of thinking of good powered walking as passive walking with a small amount of control added, I think of good powered walking, human or robotic, as highly controlled, while optimized mostly for avoiding falls and, secondarily, for minimal actuator use. However, if control is paramount, why is it that uncontrolled passive-dynamic walkers walk so much like humans? It seems that energy optimal, yet robust, control, which is perhaps a proxy for the evolutionary development of human coordination, arrives at solutions that have some features in common with passive-dynamics. Such optimized control, when well done, has most of the motor effort titrated out, yet always at the ready. Thus, highly controlled systems can deceptively-look almost “passive”.
Now, instead of thinking of good powered walking as passive walking with a small amount of control added, I think of good powered walking, human or robotic, as highly controlled, while optimized mostly for avoiding falls and, secondarily, for minimal actuator use. However, if control is paramount, why is it that uncontrolled passive-dynamic walkers walk so much like humans? It seems that energy optimal, yet robust, control, which is perhaps a proxy for the evolutionary development of human coordination, arrives at solutions that have some features in common with passive-dynamics. Such optimized control, when well done, has most of the motor effort titrated out, yet always at the ready. Thus, highly controlled systems can deceptively-look almost “passive”.
The challenge in robotics control today is the posing and solving of appropriate optimization problems. One needs an appropriate control architecture and an appropriate objective (“loss function”). The optimization problem needs to be solvable, or approximately solvable, and the model-based solution policy needs to be robust against unknown disturbances and unknown modeling errors.
The challenge in robotics control today is the posing and solving of appropriate optimization problems. One needs an appropriate control architecture and an appropriate objective (“loss function”). The optimization problem needs to be solvable, or approximately solvable, and the model-based solution policy needs to be robust against unknown disturbances and unknown modeling errors.
Speaker Bio: Prof. Andy Ruina's graduate education was in solid mechanics (in particular, friction) at Brown University. Those interests morphed into biomechanics, dynamics and robotics. Recently, he is primarily interested in the mechanics and control of underactuated motion and locomotion. His lab has made some of the most energy-effective walking robots. See ruina.org for more information.
Youtube Recording: Unfortunately, the recording for this session could not be preserved due to a technical issue.
May 21: UAV Development Cycle: Technical Benchmarks and Lessons in Innovation
May 21: UAV Development Cycle: Technical Benchmarks and Lessons in Innovation
Speaker: Dr. Tanvi Prakash (L&T Precision Engineering and Systems, India)
Speaker: Dr. Tanvi Prakash (L&T Precision Engineering and Systems, India)
Abstract: This session aims to provide participants with an understanding of the latest practices in design and development of high-performance UAVs, drawing on perspectives and challenges from ongoing programmes in the Indian industry. The session will walk participants through the stages of the UAV development cycle, starting from conceptual design and initial sizing, to detailed design and prototyping, and flight-test engineering. The talk aims to give an insight into the technical and logistical rigour involved, with a real- world focus, including some interesting case studies from L&T's UAV portfolio. The talk will show proven ways in which designs can be taken from paper to actual field conditions, by surmounting unique challenges such as terrain, weather, and operational aspects to develop a product that satisfactorily meets customer expectations. The session is targeted at professionals and students, interested in Aircraft Design, New Product Development, Prototyping, and Flight Test Engineering.
Abstract: This session aims to provide participants with an understanding of the latest practices in design and development of high-performance UAVs, drawing on perspectives and challenges from ongoing programmes in the Indian industry. The session will walk participants through the stages of the UAV development cycle, starting from conceptual design and initial sizing, to detailed design and prototyping, and flight-test engineering. The talk aims to give an insight into the technical and logistical rigour involved, with a real- world focus, including some interesting case studies from L&T's UAV portfolio. The talk will show proven ways in which designs can be taken from paper to actual field conditions, by surmounting unique challenges such as terrain, weather, and operational aspects to develop a product that satisfactorily meets customer expectations. The session is targeted at professionals and students, interested in Aircraft Design, New Product Development, Prototyping, and Flight Test Engineering.
Speaker Bio: Dr. Tanvi Prakash is an Aircraft Designer serving as Deputy General Manager (UAV) at Larsen and Toubro, a leading Indian defence contractor. She is an alumnus of VJTI Mechanical (2011) and IIT Bombay Aerospace. She has received the prestigious INAE Young Associate award and the Women in Aviation - Harvard scholarship for Emerging Leaders in Aviation. Her areas of expertise are Aircraft Design, New Product Development, and Flight Test Engineering. Her most recent UAV development assignment was leading the design and development of Teer, a jet-powered aerial target UAV, for the Indian Army. Dr. Tanvi led L&T’s team across the entire development cycle, including Design, Prototyping, Flight Testing, and Serial Production. She also serves as a UAV test pilot, conducting trials across India for various defence customers across India. Currently, she focuses on advancing innovative UAV configurations like tiltrotors, flying bullets, and morphing wings.
Youtube Recording: @ Robotics Café
May 28 (Talk-1): Not Just Where to Move, But How to Coordinate: The Challenge of Dual-Arm Manipulation
May 28 (Talk-1): Not Just Where to Move, But How to Coordinate: The Challenge of Dual-Arm Manipulation
Speaker: Debojit Das (M. Tech. Student, Indian Institute of Technology Gandhinagar)
Speaker: Debojit Das (M. Tech. Student, Indian Institute of Technology Gandhinagar)
Abstract: A robot with two arms does not automatically become twice as capable; sometimes, it becomes twice as confused. Many failures in manipulation do not happen because the robot lacks a trajectory or goal, but because it does not know how the relationship between its arms should evolve during the task, especially in the face of perturbations, or asymmetry in conditions. Through examples such as handovers, pouring, insertion, and cooperative transport, this talk looks at the hidden coordination problem behind dual-arm manipulation, and discusses how structured coordination primitives may offer a path toward more general and reliable robot behavior. The talk addresses two central questions, how a robot can use language and visual context to decide what kind of coordination is needed in the first place, and how it can execute such coordinated control reliably.
Abstract: A robot with two arms does not automatically become twice as capable; sometimes, it becomes twice as confused. Many failures in manipulation do not happen because the robot lacks a trajectory or goal, but because it does not know how the relationship between its arms should evolve during the task, especially in the face of perturbations, or asymmetry in conditions. Through examples such as handovers, pouring, insertion, and cooperative transport, this talk looks at the hidden coordination problem behind dual-arm manipulation, and discusses how structured coordination primitives may offer a path toward more general and reliable robot behavior. The talk addresses two central questions, how a robot can use language and visual context to decide what kind of coordination is needed in the first place, and how it can execute such coordinated control reliably.
Speaker Bio: Debojit Das is an early-career robotics researcher working at the intersection of bimanual manipulation, robotic grasping, control, and coordination. He is currently pursuing a B.Tech-M.Tech Dual Degree in Mechanical Engineering at the Indian Institute of Technology Gandhinagar, where he is a member of the IITGN Robotics Lab and is advised by Prof. Harish P. M. He is presently a visiting researcher at the Smart Robots Design Lab, Tohoku University, Japan, where his work focuses on developing robotic systems capable of more intelligent and coordinated real-world bimanual manipulation.
He has previously worked with the Advanced Robotics Team at Addverb Technologies, contributing to robotics research and development from an industrial perspective. Beyond research, he has also been a TEDx speaker and has spoken at ROSCon India 2025, where he shared his work and ideas with broader technical and public audiences.
More information about his work can be found at Website and Github.
Youtube Recording: @ Robotics Café
May 28 (Talk-2): Geometry-Grounded Segment Matching for Robust Robot Perception
May 28 (Talk-2): Geometry-Grounded Segment Matching for Robust Robot Perception
Speaker: Rohit Jayanti (Doctoral Student, IIIT-Hyderabad)
Speaker: Rohit Jayanti (Doctoral Student, IIIT-Hyderabad)
Abstract: Modern vision models allow robots to perceive the physical world through semantically meaningful image segments rather than fragile pixel-level keypoints. However, associating these segments across wide-baseline viewpoints, where perspective shifts and up to 180-degree rotations occur, remains a major challenge for standard 2D appearance-based encoders. This talk will discuss SegMASt3R (presented as a Spotlight at NeurIPS 2025), a framework that leverages the spatial inductive biases of 3D foundation models to achieve robust wide-baseline segment matching. This approach outperforms state-of-the-art baselines, including SAM2 and DINOv2, by up to 30% in matching accuracy in challenging settings. Furthermore, the talk will discuss how accurate segment matching translates directly to significant performance gains in key downstream robotics tasks, namely 3D Instance Mapping and Object-Relative Topological Navigation.
Abstract: Modern vision models allow robots to perceive the physical world through semantically meaningful image segments rather than fragile pixel-level keypoints. However, associating these segments across wide-baseline viewpoints, where perspective shifts and up to 180-degree rotations occur, remains a major challenge for standard 2D appearance-based encoders. This talk will discuss SegMASt3R (presented as a Spotlight at NeurIPS 2025), a framework that leverages the spatial inductive biases of 3D foundation models to achieve robust wide-baseline segment matching. This approach outperforms state-of-the-art baselines, including SAM2 and DINOv2, by up to 30% in matching accuracy in challenging settings. Furthermore, the talk will discuss how accurate segment matching translates directly to significant performance gains in key downstream robotics tasks, namely 3D Instance Mapping and Object-Relative Topological Navigation.
Speaker Bio: Rohit is a Graduate Researcher at the Robotics Research Center (RRC), IIIT-Hyderabad, advised by Prof. Madhav Krishna, where his work focuses on 3D scene understanding, geometric foundation models, and topological robot navigation. Previously, he was a founding engineer at Spatial Guide, an augmented-reality (AR) startup accelerated through the Techstars Workforce Development program (in Denver, Colorado, Cohort of 2022).
Youtube Recording: @ Robotics Café
June 2026
June 2026
June 4: Robotics at Crossroads: Is E2E Systems the Way Forward?
June 4: Robotics at Crossroads: Is E2E Systems the Way Forward?
Speaker: Prof. K Madhava Krishna (IIIT Hyderabad, India)
Speaker: Prof. K Madhava Krishna (IIIT Hyderabad, India)
Abstract: Traditional robotic systems have relied extensively on modularity, where perception and planning modules work in tango, wherein the perception stack brings together SLAM, relocalization and the navigation block comprises of planning, navigation, collision avoidance and control. While modularity brings about explainability and interpretability there are the challenges associated with propagation and amplification of errors via the modules and the challenges of modules being developed in silos where the left hand is unaware of what the right hand does. In this talk we look at how Post Modern AI and its advocacy of complete end2end systems is impacting robotics and argue how a judicious mix of classical robotics stack with the modern AI stack could still be the way to go especially where compute and data are at a premium and deployability is paramount. This argument stems though case studies we present based on our diverse experiences at the Robotics Research Center, IIIT Hyderabad. These test cases include both indoor and outdoor navigation as well as robotic manipulation.
Abstract: Traditional robotic systems have relied extensively on modularity, where perception and planning modules work in tango, wherein the perception stack brings together SLAM, relocalization and the navigation block comprises of planning, navigation, collision avoidance and control. While modularity brings about explainability and interpretability there are the challenges associated with propagation and amplification of errors via the modules and the challenges of modules being developed in silos where the left hand is unaware of what the right hand does. In this talk we look at how Post Modern AI and its advocacy of complete end2end systems is impacting robotics and argue how a judicious mix of classical robotics stack with the modern AI stack could still be the way to go especially where compute and data are at a premium and deployability is paramount. This argument stems though case studies we present based on our diverse experiences at the Robotics Research Center, IIIT Hyderabad. These test cases include both indoor and outdoor navigation as well as robotic manipulation.
Speaker Bio: Prof. K Madhava Krishna, FNAE, is the Raj Reddy Endowment Chair Professor and heads the Robotics Research Center (RRC), at IIIT Hyderabad for the last 21 years and the Kohli Center for Intelligent Systems, (KCIS) for the last 5 years. His primary research areas include Robotic Vision, Autonomous Navigation as Applied to Drones, Indoor and Outdoor Robots . RRC is nationally renowned and internationally acknowledged for its contributions to robotics research and is ranked number one in the country, top 10 in Asia and amongst top 50 in the world according to csrankings.org , where the rankings are based on publications in premier robotics forums. RRC works closely with the industry, public organizations, defense labs and the GoI on many funded projects and research collaborations on diverse set of areas ranging from robotic vision and learning to dynamics and control.
Youtube Recording: @ Robotics Café
June 11: Fast Feature Field (F³): A Predictive Representation of Events
June 11: Fast Feature Field (F³): A Predictive Representation of Events
Speaker: Prof. Pratik Chaudhari (University of Pennsylvania, USA)
Speaker: Prof. Pratik Chaudhari (University of Pennsylvania, USA)
Abstract: The human eye perceives its environment quite differently from a phone camera. While the camera creates an array of pixels---roughly 50 million of them in modern phones---the eye is more judicious. It triggers a message to the brain only when the intensity of light at a pixel changes beyond a threshold. This is why our eyes can register very fast motion and see much better in dim light or across illumination changes. “Event cameras” emulate this processing in the human eye. But since they do not really create an “image”, working with these new cameras requires rethinking a lot of fundamentals of computer vision. I will first describe a mathematical argument for building representations of event camera that retain information about structure and motion in the scene. This representation, that we call fast feature field (F³) is obtained by learning to predict future events from past events. F³ is effectively a multi-channel continuous-time video, and it can be computed at more than 120 Hz on HD data on a desktop GPU. It can be used to perform tasks such as optical flow estimation, semantic segmentation, and monocular metric depth estimation at 25—75 Hz. F³ works robustly on data from different robotic platforms, across different lighting conditions and environments. One of the biggest challenges of working with event cameras is the lack of co-registered, time-synchronized robot perception data that can be used for cross-annotations. To that end, I will also describe ongoing work on Neurosim, a high-performance library for simulating dynamic vision sensors, RGB cameras, depth sensors and inertial sensors, and agile quadrotor dynamics. It can achieve frame rates as high as 2700 FPS on a desktop GPU and 1.5 kHz closed-loop real-time control. Neurosim and its accompanying communication library called Cortex unlock downstream applications such as online self-supervised learning of neuromorphic perception and control.
Abstract: The human eye perceives its environment quite differently from a phone camera. While the camera creates an array of pixels---roughly 50 million of them in modern phones---the eye is more judicious. It triggers a message to the brain only when the intensity of light at a pixel changes beyond a threshold. This is why our eyes can register very fast motion and see much better in dim light or across illumination changes. “Event cameras” emulate this processing in the human eye. But since they do not really create an “image”, working with these new cameras requires rethinking a lot of fundamentals of computer vision. I will first describe a mathematical argument for building representations of event camera that retain information about structure and motion in the scene. This representation, that we call fast feature field (F³) is obtained by learning to predict future events from past events. F³ is effectively a multi-channel continuous-time video, and it can be computed at more than 120 Hz on HD data on a desktop GPU. It can be used to perform tasks such as optical flow estimation, semantic segmentation, and monocular metric depth estimation at 25—75 Hz. F³ works robustly on data from different robotic platforms, across different lighting conditions and environments. One of the biggest challenges of working with event cameras is the lack of co-registered, time-synchronized robot perception data that can be used for cross-annotations. To that end, I will also describe ongoing work on Neurosim, a high-performance library for simulating dynamic vision sensors, RGB cameras, depth sensors and inertial sensors, and agile quadrotor dynamics. It can achieve frame rates as high as 2700 FPS on a desktop GPU and 1.5 kHz closed-loop real-time control. Neurosim and its accompanying communication library called Cortex unlock downstream applications such as online self-supervised learning of neuromorphic perception and control.
Speaker Bio: Prof. Pratik Chaudhari is an Assistant Professor in Electrical and Systems Engineering and Computer and Information Science at the University of Pennsylvania. He is a core member of the General Robotics, Automation, Sensing and Perception (GRASP) Laboratory. From 2018-19, he was a Senior Applied Scientist at Amazon Web Services and a Postdoctoral Scholar in Computing and Mathematical Sciences at Caltech. Pratik received his PhD (2018) in Computer Science from UCLA, and his Master's (2012) and Engineer's (2014) degrees in Aeronautics and Astronautics from MIT. He was a part of NuTonomy Inc. (now Hyundai-Aptiv Motional) from 2014-16. He is the recipient of the Amazon Machine Learning Research Award (2020), NSF CAREER award (2022) and the Intel Rising Star Faculty Award (2022).
Youtube Recording: @ Robotics Café
June 18: From Simulation to Sky: Developing Autonomous UAVs with MATLAB
June 18: From Simulation to Sky: Developing Autonomous UAVs with MATLAB
Speaker: Dr. N Dhananjay (MathWorks India Pvt. Ltd., Bengaluru, India)
Speaker: Dr. N Dhananjay (MathWorks India Pvt. Ltd., Bengaluru, India)
Abstract: Developing autonomous UAVs requires navigating a broad engineering stack — vehicle dynamics, sensor modeling, path planning, mission management, and real-time communication — while managing the risk and cost of hardware testing. This talk presents the MATLAB UAV Toolbox as an end-to-end framework for taking a UAV algorithm from first-principles simulation to real flight, covering both fixed-wing and multirotor vehicle types. We walk you through the toolbox's layered architecture: reduced-order guidance models (6-state fixed-wing, 12-state multirotor) that support automatic code generation; a scenario simulation environment with LiDAR, GPS/INS, radar, and atmospheric turbulence models; and a path planning suite spanning 3D Dubins curves, waypoint following, coverage planning, and Vector Field Histogram obstacle avoidance. We then show how Simulink can be used to enable model-based design workflows. We highlight software-in-the-loop (SITL), hardware-in-the-loop (HITL), and deployment to PX4 and ArduPilot autopilots via auto-generated C code. We discuss the MAVLink communication layer, including a microservice pattern that abstracts parameter management, command dispatch, and mission upload behind high-level MATLAB classes. We present post-flight analysis workflows — parsing ArduPilot and PX4 logs into MATLAB timetables, system identification techniques for refining aerodynamic and wind turbulence models from flight data. The talk showcases important UAV workflows of interest for both academia and industry.
Abstract: Developing autonomous UAVs requires navigating a broad engineering stack — vehicle dynamics, sensor modeling, path planning, mission management, and real-time communication — while managing the risk and cost of hardware testing. This talk presents the MATLAB UAV Toolbox as an end-to-end framework for taking a UAV algorithm from first-principles simulation to real flight, covering both fixed-wing and multirotor vehicle types. We walk you through the toolbox's layered architecture: reduced-order guidance models (6-state fixed-wing, 12-state multirotor) that support automatic code generation; a scenario simulation environment with LiDAR, GPS/INS, radar, and atmospheric turbulence models; and a path planning suite spanning 3D Dubins curves, waypoint following, coverage planning, and Vector Field Histogram obstacle avoidance. We then show how Simulink can be used to enable model-based design workflows. We highlight software-in-the-loop (SITL), hardware-in-the-loop (HITL), and deployment to PX4 and ArduPilot autopilots via auto-generated C code. We discuss the MAVLink communication layer, including a microservice pattern that abstracts parameter management, command dispatch, and mission upload behind high-level MATLAB classes. We present post-flight analysis workflows — parsing ArduPilot and PX4 logs into MATLAB timetables, system identification techniques for refining aerodynamic and wind turbulence models from flight data. The talk showcases important UAV workflows of interest for both academia and industry.
Speaker Bio: Dr. N Dhananjay is a Senior Software Engineer at MathWorks, India. He is a member of the Autonomous Aerial Systems Group. Prior to joining MathWorks in 2017, he worked as an Image Processing Engineer at UURMI System Pvt. Ltd., Hyderabad, India where he worked on the Autonomous Landing of UAV project. Dhananjay received his PhD in Aerospace Engineering from IISc, Bangalore, and his Master's in Electrical Engineering also from IISc. Currently he is working on developing tools in MATLAB/SIMULINK for autonomous UAV applications.
Youtube Recording: @ Robotics Café
June 25 (Talk 1): DOME: Distributed Online Learning-based Multi-Estimate Fusion for Cooperative Predictive Target Tracking Using a Robotic Swarm
June 25 (Talk 1): DOME: Distributed Online Learning-based Multi-Estimate Fusion for Cooperative Predictive Target Tracking Using a Robotic Swarm
Speaker: Shubhankar Gupta (IISc Bengaluru, India)
Speaker: Shubhankar Gupta (IISc Bengaluru, India)
Abstract: Cooperative predictive target tracking is a critical capability for robotic swarms, but real-world deployments are often plagued by high prediction bias and communication uncertainty. In this talk, we investigate the challenge of target tracking when robots interact over a randomly time-varying communication network and exhibit significant heterogeneity in their onboard sensors and prediction algorithms. To address these challenges, I will introduce the Distributed Online learning-based Multi-Estimate (DOME) fusion algorithm. DOME performs a collaborative, weighted fusion of local and socially shared predictions, adapting fusion weights online using direct feedback from prediction loss rather than relying on traditional covariance information. I will present the theoretical foundations of the algorithm and establish the convergence of the conditional expectations of the fusion weights under reasonable assumptions. Finally, I will demonstrate DOME's efficacy through comprehensive performance and scalability tests, showing that DOME achieves up to 74% lower prediction loss compared to covariance-based and online learning-based decentralized fusion baselines, particularly in environments with significant model drift and unreliable communication.
Abstract: Cooperative predictive target tracking is a critical capability for robotic swarms, but real-world deployments are often plagued by high prediction bias and communication uncertainty. In this talk, we investigate the challenge of target tracking when robots interact over a randomly time-varying communication network and exhibit significant heterogeneity in their onboard sensors and prediction algorithms. To address these challenges, I will introduce the Distributed Online learning-based Multi-Estimate (DOME) fusion algorithm. DOME performs a collaborative, weighted fusion of local and socially shared predictions, adapting fusion weights online using direct feedback from prediction loss rather than relying on traditional covariance information. I will present the theoretical foundations of the algorithm and establish the convergence of the conditional expectations of the fusion weights under reasonable assumptions. Finally, I will demonstrate DOME's efficacy through comprehensive performance and scalability tests, showing that DOME achieves up to 74% lower prediction loss compared to covariance-based and online learning-based decentralized fusion baselines, particularly in environments with significant model drift and unreliable communication.
Speaker Bio: Shubhankar Gupta is a Ph.D. candidate in the Department of Aerospace Engineering at the Indian Institute of Science (IISc), Bengaluru, India. His research in the AI and Robotics Lab (AIRL) focuses on distributed online learning, swarm robotics, multi-agent systems, decentralized optimization, and control theory. Supported by the 2024 KIAC Ph.D. Fellowship, his work has appeared in venues like the Transactions on Machine Learning Research (TMLR) and the AAAI Conference on Artificial Intelligence (AAAI). He received an M.S. in Aeronautics and Astronautics from Purdue University (West Lafayette, IN, US) and a B.Tech. in Mechanical Engineering from the Indian Institute of Technology Kanpur (IITK), and previously worked as a Project Assistant at the National Aerospace Laboratories (NAL) and the Aeronautical Development Agency (ADA) in Bengaluru, India.
Youtube Recording: @ Robotics Café
June 25 (Talk 2): Multi-Agent Patrolling: Questions, Insights, and Contributions
June 25 (Talk 2): Multi-Agent Patrolling: Questions, Insights, and Contributions
Speaker: S Deepak Mallya (IIT Bombay)
Speaker: S Deepak Mallya (IIT Bombay)
Abstract: Multi-agent patrolling is a fundamental problem in robotics with applications ranging from campus and infrastructure surveillance to environmental monitoring and autonomous security systems. The objective is deceptively simple: coordinate a team of agents to repeatedly visit important locations in an environment. Yet, the problem gives rise to rich questions involving optimization, graph theory, task allocation, and long-term autonomous decision-making. This talk introduces the multi-agent patrolling problem, discusses the diverse formulations studied in the literature, and reviews the key approaches that have been proposed over the past two decades. Particular attention is given to several intriguing open questions concerning the existence, structure, and optimality of patrol strategies. The talk then presents research contributions that address some of these questions through a unified framework based on recurrent patrol strategies. The discussion covers both theoretical results on the nature of near-optimal solutions and practical algorithms for priority-based patrolling with multiple agents, highlighting how these results contribute to practical autonomous surveillance systems.
Abstract: Multi-agent patrolling is a fundamental problem in robotics with applications ranging from campus and infrastructure surveillance to environmental monitoring and autonomous security systems. The objective is deceptively simple: coordinate a team of agents to repeatedly visit important locations in an environment. Yet, the problem gives rise to rich questions involving optimization, graph theory, task allocation, and long-term autonomous decision-making. This talk introduces the multi-agent patrolling problem, discusses the diverse formulations studied in the literature, and reviews the key approaches that have been proposed over the past two decades. Particular attention is given to several intriguing open questions concerning the existence, structure, and optimality of patrol strategies. The talk then presents research contributions that address some of these questions through a unified framework based on recurrent patrol strategies. The discussion covers both theoretical results on the nature of near-optimal solutions and practical algorithms for priority-based patrolling with multiple agents, highlighting how these results contribute to practical autonomous surveillance systems.
Speaker Bio: Deepak Mallya is a PhD in Systems and Control currently working at dotRobotics, where he is leading efforts to productize an existing robotics prototype into a scalable and deployment-ready product. His professional interests include multi-agent systems, robotics, autonomous systems, and distributed intelligence. Beyond his technical work, he is deeply interested in Indic knowledge systems and Hindustani classical music. His work reflects a blend of rigorous engineering, product thinking, and a continuing engagement with India's intellectual and cultural traditions.
Youtube Recording: @ Robotics Café
July 2026
July 2026
July 2: From Theory to Practice: Bringing Robots into the Real World
July 2: From Theory to Practice: Bringing Robots into the Real World
Speaker: Dr. Vighnesh Vatsal (TCS Research, India)
Speaker: Dr. Vighnesh Vatsal (TCS Research, India)
Abstract: Robotics is hard. Not because we have too many gaps in our theories. But because it tries to bridge the gap between the computational world and our physical world-- going from bits to atoms. Our models and simulators cannot capture reality in all its complexities, so roboticists are always working partially in the dark. In this talk, we will explore how robotics can be applied to real-world scenarios even with these inherent shortcomings-- how we optimize for things under our control, and account for those that we cannot. As examples, we will look at industrial mobile manipulation, human-robot interaction in the form of wearable devices, humanoid robots, and space robotics. After outlining the key technologies and practices used in these domains, we will discuss the path forward for robotics as a multidisciplinary field, and how you can contribute to it regardless of your specific area of expertise.
Abstract: Robotics is hard. Not because we have too many gaps in our theories. But because it tries to bridge the gap between the computational world and our physical world-- going from bits to atoms. Our models and simulators cannot capture reality in all its complexities, so roboticists are always working partially in the dark. In this talk, we will explore how robotics can be applied to real-world scenarios even with these inherent shortcomings-- how we optimize for things under our control, and account for those that we cannot. As examples, we will look at industrial mobile manipulation, human-robot interaction in the form of wearable devices, humanoid robots, and space robotics. After outlining the key technologies and practices used in these domains, we will discuss the path forward for robotics as a multidisciplinary field, and how you can contribute to it regardless of your specific area of expertise.
Speaker Bio: Dr. Vighnesh Vatsal is a Scientist at TCS Research, Tata Consultancy Services, working in the areas of robotics, AI and control. He received a Ph.D. from Cornell University in 2020 in Mechanical Engineering with a focus on Human-Robot Collaboration. He previously received B.Tech. and M.Tech. degrees in Mechanical Engineering from IIT Bombay with a focus on automation and control. His research interests span the full spectrum of robotics and automation - including mobile manipulation, grasping, reinforcement learning and applied machine learning, exoskeletons, and space robotics.
Youtube Recording: @ Robotics Café
July 9: Human-Robot Interaction in Wearable Systems: Comfort, Adaptation, and Immersion
July 9: Human-Robot Interaction in Wearable Systems: Comfort, Adaptation, and Immersion
Speaker: Prof. Vineet Vashista (IIT Gandhinagar, India)
Speaker: Prof. Vineet Vashista (IIT Gandhinagar, India)
Abstract: Exoskeletons and exosuits, designed to assist or resist human movement, play a vital role in both rehabilitation and performance augmentation. Their effectiveness depends on transparent and comfortable human-robot interaction, where user adaptation to device dynamics is central. This talk presents insights from gait adaptation studies using cable-driven exoskeletons and passive exosuits, highlighting how design choices such as cable routing and device-to-body forces influence outcomes. It also explores immersive interventions that integrate virtual reality with wearable systems to enhance intent detection and user engagement. Ongoing work at the Human-Centered Robotics Lab at IIT Gandhinagar demonstrates how multidisciplinary approaches can advance wearable robotics to restore mobility and augment human performance.
Abstract: Exoskeletons and exosuits, designed to assist or resist human movement, play a vital role in both rehabilitation and performance augmentation. Their effectiveness depends on transparent and comfortable human-robot interaction, where user adaptation to device dynamics is central. This talk presents insights from gait adaptation studies using cable-driven exoskeletons and passive exosuits, highlighting how design choices such as cable routing and device-to-body forces influence outcomes. It also explores immersive interventions that integrate virtual reality with wearable systems to enhance intent detection and user engagement. Ongoing work at the Human-Centered Robotics Lab at IIT Gandhinagar demonstrates how multidisciplinary approaches can advance wearable robotics to restore mobility and augment human performance.
Speaker Bio: Dr. Vineet Vashista is an Associate Professor at the Indian Institute of Technology Gandhinagar, where he leads the Human-Centered Robotics (HCR) Laboratory. He earned his Ph.D. in Mechanical Engineering from Columbia University in 2015 and a Master’s in Mechanical Design from IIT Delhi in 2008. His research focuses on human-centered robotic technologies inspired by human motor adaptation, with applications in rehabilitation, wearable robotics, and performance augmentation. Supported by major national and international funding agencies, his work investigates neuro-musculoskeletal processes to design devices that measure and influence human movement. Dr. Vashista is a recipient of the Haryana Yuva Vigyan Ratna Award (2022) and a member of INYAS (2023).
Youtube Recording: @ Robotics Café
July 16: From Pixels to Planning: World Representations as the Lever for Robot Autonomy
July 16: From Pixels to Planning: World Representations as the Lever for Robot Autonomy
Speaker: Prof. Karthik Dantu (University at Buffalo, USA)
Speaker: Prof. Karthik Dantu (University at Buffalo, USA)
Abstract: Robot planning has long relied on maps that are easy to construct but hard to reason over at scale: dense grids, raster cost maps, and fixed-resolution spaces that treat every part of the world as equally important. This talk argues that the right world representation is the true lever for scalable autonomy. Across three recent works, we show how robots can plan faster, adapt to changing conditions, and reason over richer environments by replacing uniform discretizations with structured representations that preserve the geometry, semantics, topology, and dynamics that matter for action.
Abstract: Robot planning has long relied on maps that are easy to construct but hard to reason over at scale: dense grids, raster cost maps, and fixed-resolution spaces that treat every part of the world as equally important. This talk argues that the right world representation is the true lever for scalable autonomy. Across three recent works, we show how robots can plan faster, adapt to changing conditions, and reason over richer environments by replacing uniform discretizations with structured representations that preserve the geometry, semantics, topology, and dynamics that matter for action.
CLEAR builds reusable semantic–geometric terrain abstractions from landcover and elevation, producing convex region graphs for large-scale off-road planning. HOPHY extends this idea into hierarchical hypergraphs, enabling kilometer-scale mission planning and fast replanning under weather, terrain, and agent-specific constraints. GoLF generalizes the principle to adaptive graphs over spatial fields, concentrating representational capacity near informative structure such as obstacle boundaries, narrow passages, and dynamic pedestrian context while remaining sparse in free space. Together, these works present world representations not as passive maps, but as computational fulcrums: they trade pixels for structure, recomputation for reusable abstraction, and brittle uniformity for adaptive reasoning. The result is a planning stack that moves from indoor navigation and social robot motion to large-scale off-road mission planning—faster, with comparable path quality, and with representations built for repeated decision-making in dynamic worlds.
CLEAR builds reusable semantic–geometric terrain abstractions from landcover and elevation, producing convex region graphs for large-scale off-road planning. HOPHY extends this idea into hierarchical hypergraphs, enabling kilometer-scale mission planning and fast replanning under weather, terrain, and agent-specific constraints. GoLF generalizes the principle to adaptive graphs over spatial fields, concentrating representational capacity near informative structure such as obstacle boundaries, narrow passages, and dynamic pedestrian context while remaining sparse in free space. Together, these works present world representations not as passive maps, but as computational fulcrums: they trade pixels for structure, recomputation for reusable abstraction, and brittle uniformity for adaptive reasoning. The result is a planning stack that moves from indoor navigation and social robot motion to large-scale off-road mission planning—faster, with comparable path quality, and with representations built for repeated decision-making in dynamic worlds.
Speaker Bio: Dr. Karthik Dantu is an Associate Professor in Computer Science Engineering at University at Buffalo, the Founding Director of the Center for Embodied Autonomy and Robotics, a university-wide research center in Robotics, and a Co-Director of the Masters program in Robotics. His research interests are primarily in perception and representation for robot autonomy. Karthik received his graduate degrees from University of Southern California and was a postdoctoral fellow at Harvard from 2010-2013. His work has been recognized with an NSF CAREER award in 2019, award papers at IEEE MRS 2019, ACM EICS 2019 and IEEE IPSN 2012. Beyond robotics, Karthik is a big fan of cricket, American football, good coffee and beer.
Youtube Recording: @ Robotics Café
July 23: 10s/100s/1000s/1,000,000s - Integrating Robotics, Computer Vision and Artificial Intelligence for Robust Grasping and Manipulation in Real-World Scenarios
July 23: 10s/100s/1000s/1,000,000s - Integrating Robotics, Computer Vision and Artificial Intelligence for Robust Grasping and Manipulation in Real-World Scenarios
Speaker: Dr. Juxi Leitner (Amazon, Germany)
Speaker: Dr. Juxi Leitner (Amazon, Germany)
Abstract: I will be talking about robotic manipulation and, in particular, the differences when looking at it from the lens of academic research vs industrial deployments. Robust and rigorous robotics is at the core of getting robots "out there", yet it is not always clear how this is best investigated in academic research.
Abstract: I will be talking about robotic manipulation and, in particular, the differences when looking at it from the lens of academic research vs industrial deployments. Robust and rigorous robotics is at the core of getting robots "out there", yet it is not always clear how this is best investigated in academic research.
Speaker Bio: Dr. Juxi Leitner is an Applied Science Manager at Amazon in Berlin, Germany. He has spent the better part of the last two decades conducting robotics research across academia, government, and industry, most recently as Sr Lecturer at Monash University in Melbourne. Before joining Amazon, Juxi was founder and CTO of LYRO Robotics, an Australian, VC-backed startup creating general-purpose pick-and-place robots and the software to deploy them in the food supply chain. In 2017, his team from the Australian Centre for Robotic Vision won the Amazon Robotics Challenge.
Youtube Recording: @ Robotics Café
July 30 (Talk 1): Nonlinear Guidance for Impact Angle Constrained Target Interception Under FOV and Input Constraints
July 30 (Talk 1): Nonlinear Guidance for Impact Angle Constrained Target Interception Under FOV and Input Constraints
Speaker: Sayantan Pal (IIT Kharagpur)
Speaker: Sayantan Pal (IIT Kharagpur)
Abstract: Modern interceptor systems must achieve precise terminal impact angles while satisfying practical constraints arising from seeker field-of-view (FOV) limitations and actuator saturation. In this talk, I will present a nonlinear guidance framework for intercepting nonmaneuvering moving targets under simultaneous impact-angle, FOV, and input constraints. The proposed approach employs an Asymmetric Integral Barrier Lyapunov Function (AIBLF) to enforce the FOV constraint and introduces an asymmetric smooth bounded function for virtual control design and the modeling of input saturation. A first-order auxiliary filter is incorporated to compensate for saturation effects, resulting in a continuous guidance law with guaranteed stability. I will discuss the theoretical development of the guidance law and demonstrate its performance through simulation studies and comparisons with existing methods, highlighting improved control efficiency while satisfying all prescribed constraints.
Abstract: Modern interceptor systems must achieve precise terminal impact angles while satisfying practical constraints arising from seeker field-of-view (FOV) limitations and actuator saturation. In this talk, I will present a nonlinear guidance framework for intercepting nonmaneuvering moving targets under simultaneous impact-angle, FOV, and input constraints. The proposed approach employs an Asymmetric Integral Barrier Lyapunov Function (AIBLF) to enforce the FOV constraint and introduces an asymmetric smooth bounded function for virtual control design and the modeling of input saturation. A first-order auxiliary filter is incorporated to compensate for saturation effects, resulting in a continuous guidance law with guaranteed stability. I will discuss the theoretical development of the guidance law and demonstrate its performance through simulation studies and comparisons with existing methods, highlighting improved control efficiency while satisfying all prescribed constraints.
Speaker Bio: Sayantan Pal received the B.Tech. degree in Mechanical Engineering from Maulana Abul Kalam Azad University of Technology, Kolkata, India, in 2016, and the M.Tech. degree in Mechatronics from the Indian Institute of Engineering Science and Technology (IIEST), Shibpur, India, in 2020. He is currently pursuing the Ph.D. degree in Aerospace Engineering at the Indian Institute of Technology Kharagpur, India. From 2020 to 2021, he was a Project Associate at CSIR–Central Mechanical Engineering Research Institute, Durgapur, India. His research interests include path planning, guidance and control of autonomous vehicles, and optimization.
Youtube Recording: @ Robotics Café
July 30 (Talk 2): Learning to Recover from Plan Execution Errors during Robot Manipulation: A Neuro-symbolic Approach
July 30 (Talk 2): Learning to Recover from Plan Execution Errors during Robot Manipulation: A Neuro-symbolic Approach
Speaker: Namasivayam K (IIT Delhi)
Speaker: Namasivayam K (IIT Delhi)
Abstract: Foundation models, through large-scale pre-training, have become highly capable at visuo-lingual understanding, commonsense reasoning, and instruction following. However, they continue to struggle with long-horizon planning, particularly when tasks require inductive or grounded physical reasoning. Symbolic approaches provide a complementary strength in such settings, suggesting that integrating these two paradigms can lead to more capable autonomous systems.
Abstract: Foundation models, through large-scale pre-training, have become highly capable at visuo-lingual understanding, commonsense reasoning, and instruction following. However, they continue to struggle with long-horizon planning, particularly when tasks require inductive or grounded physical reasoning. Symbolic approaches provide a complementary strength in such settings, suggesting that integrating these two paradigms can lead to more capable autonomous systems.
The cognitive model of the human brain, which distinguishes between fast and slow systems, provides a useful inspiration for this integration. Neural models are well suited for fast, reflexive tasks such as object detection and language understanding, whereas symbolic methods are better equipped for deliberative planning and complex reasoning. Motivated by this perspective, my research combines the representational power of large neural models with the reasoning capabilities of classical symbolic approaches, with the goal of developing autonomous systems that are both more interpretable and more verifiable in their interactions with humans.
The cognitive model of the human brain, which distinguishes between fast and slow systems, provides a useful inspiration for this integration. Neural models are well suited for fast, reflexive tasks such as object detection and language understanding, whereas symbolic methods are better equipped for deliberative planning and complex reasoning. Motivated by this perspective, my research combines the representational power of large neural models with the reasoning capabilities of classical symbolic approaches, with the goal of developing autonomous systems that are both more interpretable and more verifiable in their interactions with humans.
This talk demonstrates this idea through a concrete problem: automatically detecting and recovering from failures during long-horizon manipulation tasks, such as building a tower, without requiring any annotated failure data. I will show how a neural state representation, a learned transition function, and a failure discriminator work together with symbolic search to detect and localise failures, and subsequently re-plan using a learned heuristic distance function.
This talk demonstrates this idea through a concrete problem: automatically detecting and recovering from failures during long-horizon manipulation tasks, such as building a tower, without requiring any annotated failure data. I will show how a neural state representation, a learned transition function, and a failure discriminator work together with symbolic search to detect and localise failures, and subsequently re-plan using a learned heuristic distance function.
Speaker Bio: Namasivayam Kalithasan is a PhD candidate at IIT Delhi, where he is advised by Prof. Rohan Paul and Prof. Parag Singla. He is a recipient of the Google PhD Fellowship. His research focuses on neuro-symbolic AI for embodied intelligence, spanning robot manipulation, verifiable and trustworthy robotics, and human-centric decision-making. Before joining IIT Delhi, he completed his MSc in Mathematics and Computing at IIT Hyderabad. Alongside his research, he is passionate about teaching and has received the Outstanding TA Award at IIT Delhi on multiple occasions for his contributions to the Introductory AI and graduate-level Autonomous Systems courses.
Youtube Recording: @ Robotics Café
August 2026
August 2026
August 6: Building and flying unconventional UAVs
August 6: Building and flying unconventional UAVs
Speaker: Prof. Harikumar Kandath (IIIT Hyderabad, India)
Speaker: Prof. Harikumar Kandath (IIIT Hyderabad, India)
Abstract: This discussion is about two unconventional Uncrewed Aerial Vehicles (UAVs) that are designed and flight tested at my research lab here at RRC, IIIT Hyderabad. The first one is a hybrid UAV that can fly like a flapping-wing UAV and also like a fixed-wing UAV. The hybrid UAV has an enhanced flight regime when compared to a fixed-wing UAV and a higher range and flight speed when compared to a flapping UAV. The second one is a spinning quadrotor with additional lifting surfaces. They can provide additional upward thrust while take-off and hence suitable for applications where climbing is the primary manoeuvre like inspection of tall industrial structures. Basic design approaches, issues in modeling and control, and flight testing for these two vehicles will be discussed here.
Abstract: This discussion is about two unconventional Uncrewed Aerial Vehicles (UAVs) that are designed and flight tested at my research lab here at RRC, IIIT Hyderabad. The first one is a hybrid UAV that can fly like a flapping-wing UAV and also like a fixed-wing UAV. The hybrid UAV has an enhanced flight regime when compared to a fixed-wing UAV and a higher range and flight speed when compared to a flapping UAV. The second one is a spinning quadrotor with additional lifting surfaces. They can provide additional upward thrust while take-off and hence suitable for applications where climbing is the primary manoeuvre like inspection of tall industrial structures. Basic design approaches, issues in modeling and control, and flight testing for these two vehicles will be discussed here.
Speaker Bio: Dr. Harikumar Kandath received the Ph.D. degree in aerospace engineering from the Indian Institute of Science (IISc), Bengaluru, India, in 2015. He has held research positions at NTU Singapore, SUTD Singapore and University of Macau, China. He is working as an Assistant Professor with the Robotics Research Centre, International Institute of Information Technology, Hyderabad, India since March 2020. His research interests are applications of control theory to unmanned systems, UAV Design and flight dynamics.
Youtube Recording: @ Robotics Café
August 13: Field and Algorithmic Robotics
August 13: Field and Algorithmic Robotics
Speaker: Dr. Nare Karapetyan (Woods Hole Oceanographic Institute (WHOI), USA)
Speaker: Dr. Nare Karapetyan (Woods Hole Oceanographic Institute (WHOI), USA)
Abstract: Efforts to understand our world drive the development of autonomous systems capable of effectively planning exploration tasks across applications like scientific sampling, environmental monitoring, surveillance, and search and rescue. When deployed in real-world settings, these robotic systems face dynamic changes and environmental uncertainties that dramatically increase decision complexity, making planning challenging. By leveraging the geometric properties of target areas, planning can be reframed as a combinatorial optimization problem, reducing complexity and enabling the breakdown of tasks into manageable subproblems. This talk presents a hierarchical approach to creating robust exploration and coverage plans, focusing on (i) generating global trajectory plans and (ii) adapting these trajectories in response to dynamic changes. We will discuss single and multi-robot coverage strategies that consider obstacles, environmental features, and sensor-specific data collection optimization, along with approaches to incorporate uncertainties into these plans. Real-world applications, such as automated scientific sampling in marine environments, will demonstrate the feasibility and impact of these techniques.
Abstract: Efforts to understand our world drive the development of autonomous systems capable of effectively planning exploration tasks across applications like scientific sampling, environmental monitoring, surveillance, and search and rescue. When deployed in real-world settings, these robotic systems face dynamic changes and environmental uncertainties that dramatically increase decision complexity, making planning challenging. By leveraging the geometric properties of target areas, planning can be reframed as a combinatorial optimization problem, reducing complexity and enabling the breakdown of tasks into manageable subproblems. This talk presents a hierarchical approach to creating robust exploration and coverage plans, focusing on (i) generating global trajectory plans and (ii) adapting these trajectories in response to dynamic changes. We will discuss single and multi-robot coverage strategies that consider obstacles, environmental features, and sensor-specific data collection optimization, along with approaches to incorporate uncertainties into these plans. Real-world applications, such as automated scientific sampling in marine environments, will demonstrate the feasibility and impact of these techniques.
Speaker Bio: Dr. Nare Karapetyan is a Tenure Track Assistant Scientist and an MIT-WHOI JP faculty at Woods Hole Oceanographic Institute (WHOI). Previously was a postdoctoral associate at Maryland Robotics Center at University of Maryland (UMD) where she was working with Dr. Dinesh Manocha, Dr. Pratap Tokekar and Dr. Miao Yu. She received her PhD in Computer Science from the University of South Carolina (UofSC), where she worked in the Autonomous Field Robotics Laboratory (AFRL) under the supervision of Dr. Ioannis Rekleitis. Since 2022, she is serving as an Associate Editor (AE) for the RA-L, IROS, ICRA and AURO Editorial Boards. Her research focuses on planning and exploration problems with heterogeneous multi-agent systems, with particular applications in the aquatic domain. She aims to develop more efficient, task-oriented exploration techniques for environmental monitoring and survey operations.
Youtube Recording: Not available.
August 20: Chakravyuha: Myth, Math, and Magic
August 20: Chakravyuha: Myth, Math, and Magic
Speaker: Prof. Arijit Sen (IISER Bhopal, India)
Speaker: Prof. Arijit Sen (IISER Bhopal, India)
Abstract: A group of unmanned aerial vehicles (UAVs) can efficiently obtain the information of an unknown target or location of interest. Efficacy of localization gets enhanced if UAVs form an optimal formation geometry around the target. This work presents a unified scheme of cooperative localization and tracking for a group of fixed-wing unmanned aerial vehicles. The sensors having different types and qualities are mounted on each UAV to form a heterogeneous sensing network. These UAVs are used to localize a moving target with unknown states. The performance of an unbiased estimator can be quantified by the Cramer–Rao lower bound, which is the same as the inverse of Fisher information. In this work, the conditions of optimal UAV formations are derived to maximize the Fisher information of the unknown target. Next, a generalized protocol is presented to result in the desired formations around the moving target that maximizes the information of the target. A cooperative localization algorithm is used to localize the unknown target states using the noisy measurements obtained from the heterogeneous sensing network. The efficacy of the proposed unified framework is illustrated with numerical examples.
Abstract: A group of unmanned aerial vehicles (UAVs) can efficiently obtain the information of an unknown target or location of interest. Efficacy of localization gets enhanced if UAVs form an optimal formation geometry around the target. This work presents a unified scheme of cooperative localization and tracking for a group of fixed-wing unmanned aerial vehicles. The sensors having different types and qualities are mounted on each UAV to form a heterogeneous sensing network. These UAVs are used to localize a moving target with unknown states. The performance of an unbiased estimator can be quantified by the Cramer–Rao lower bound, which is the same as the inverse of Fisher information. In this work, the conditions of optimal UAV formations are derived to maximize the Fisher information of the unknown target. Next, a generalized protocol is presented to result in the desired formations around the moving target that maximizes the information of the target. A cooperative localization algorithm is used to localize the unknown target states using the noisy measurements obtained from the heterogeneous sensing network. The efficacy of the proposed unified framework is illustrated with numerical examples.
Speaker Bio: Dr. Arijit Sen is a faculty member and researcher with a passion for control systems and engineering. His academic journey began at Jadavpur University, Kolkata, where he received his Bachelor of Engineering in Electrical Engineering in 2014. He then pursued his Master's and Ph.D. in Electrical Engineering at the Indian Institute of Technology Kanpur, graduating in 2016 and 2020 respectively. Following his Ph.D., Dr. Sen worked as a research associate at IIT Kanpur from January 2021 to June 2022. Currently, he holds a faculty position in the Department of Electrical Engineering and Computer Science at the Indian Institute of Science Education and Research Bhopal (IISERB). Dr. Sen's research interests lie broadly within the field of control engineering, encompassing areas like optimization, multi-agent systems, and robotics. His accomplishments are recognized by young author awards received from prestigious organizations: the Control Society of India (2017), the American Automatic Control Council (2018), the European Control Association (2019), and the International Federation of Automatic Control (2020).
Youtube Recording: @ Robotics Café
August 27 (Talk 1): Kernel-Based Safe Exploration in Deep Reinforcement Learning
August 27 (Talk 1): Kernel-Based Safe Exploration in Deep Reinforcement Learning
Speaker: Dr. Nikhil Singh (MPI-SWS, Germany)
Speaker: Dr. Nikhil Singh (MPI-SWS, Germany)
Abstract: Safety has been a major concern when deploying deep reinforcement learning algorithms in the real world. A promising direction that ensures that the learned policy does not visit unsafe regions is to learn a barrier function along with the policy. A barrier is a function from states to reals that assigns low values to the initial states, high values to the unsafe states, and decreases in expectation on each transition; such a function can be used to bound the probability of reaching unsafe states. Previous attempts learned a barrier function directly from exploration data, but this required either large amounts of data or restrictions on the system dynamics. In this paper, we show how kernel embeddings can be used to learn barrier functions during deep reinforcement learning for stochastic systems with unknown dynamics. Our algorithm, kernel-based safe exploration (KBSE), learns an optimal policy and a barrier simultaneously during exploration. The barriers are computed iteratively, represented as conditional mean embeddings, and provide better probabilistic safety guarantees with more exploration. The exploration algorithm uses the learned barrier functions to identify safety violations. In the case of violation, it intervenes to modify the unsafe action to a safe action, thereby ensuring that the exploration is restricted to actions that bound the probability of reaching unsafe states. We evaluate KBSE on several complex continuous control benchmarks. Experimental results establish our new algorithm to be suitable for synthesizing control policies that are probabilistically safe without degradation in reward accumulation.
Abstract: Safety has been a major concern when deploying deep reinforcement learning algorithms in the real world. A promising direction that ensures that the learned policy does not visit unsafe regions is to learn a barrier function along with the policy. A barrier is a function from states to reals that assigns low values to the initial states, high values to the unsafe states, and decreases in expectation on each transition; such a function can be used to bound the probability of reaching unsafe states. Previous attempts learned a barrier function directly from exploration data, but this required either large amounts of data or restrictions on the system dynamics. In this paper, we show how kernel embeddings can be used to learn barrier functions during deep reinforcement learning for stochastic systems with unknown dynamics. Our algorithm, kernel-based safe exploration (KBSE), learns an optimal policy and a barrier simultaneously during exploration. The barriers are computed iteratively, represented as conditional mean embeddings, and provide better probabilistic safety guarantees with more exploration. The exploration algorithm uses the learned barrier functions to identify safety violations. In the case of violation, it intervenes to modify the unsafe action to a safe action, thereby ensuring that the exploration is restricted to actions that bound the probability of reaching unsafe states. We evaluate KBSE on several complex continuous control benchmarks. Experimental results establish our new algorithm to be suitable for synthesizing control policies that are probabilistically safe without degradation in reward accumulation.
Speaker Bio: Dr. Nikhil Singh is a Postdoctoral Researcher at the Max Planck Institute for Software Systems (MPI-SWS), Germany, working with Prof. Sadegh Soudjani and Prof. Rupak Majumdar. He is currently part of the ERC-funded AutoCyPher project. He completed his PhD in Computer Science and Engineering at the Indian Institute of Technology Kanpur under Prof. Indranil Saha, supported by the Prime Minister's Research Fellowship. His research lies at the intersection of formal methods, reinforcement learning, and control theory, focusing on the design of safe, robust, and efficient learning-based controllers for autonomous and cyber-physical systems.
Youtube Recording: TBA
August 27 (Talk 2): How Do You Make a Learned Controller Safe? Two Answers from Offline Data
August 27 (Talk 2): How Do You Make a Learned Controller Safe? Two Answers from Offline Data
Speaker: Dr. Manan Tayal (Emergence AI)
Speaker: Dr. Manan Tayal (Emergence AI)
Abstract: A controller trained on data can be very good on average and still be unacceptable, because averages are the wrong thing to measure when a single collision ends the episode. Most safe reinforcement learning methods bound expected cost, which permits exactly that failure. The usual fix is to try things out in the world and learn from what goes wrong, but that is not available when the failures are the ones you were trying to avoid. So the question becomes: given a fixed dataset, no further interaction, and often no model of how the system moves, what can you actually certify? I will start from the control-theoretic answer, which is to build a function over the state space that tells you which states are still recoverable, and then keep the system inside that region. The talk covers two ways of learning such a function from static data. The first keeps it separate from the controller. V-OCBF learns a barrier value function from demonstrations and uses it as a filter that sits in front of any nominal policy and corrects actions only when needed. The second folds it into the policy itself. SafeFQL learns a safety critic that gates policy training directly, which removes the need for a dynamics model and makes deployment a single forward pass. Both are evaluated on the same navigation and locomotion benchmarks, so the comparison is concrete. I will use it to draw out which sources of error actually matter in practice, what each design costs at deployment, and where the empirical safety numbers still run ahead of anything we can prove.
Abstract: A controller trained on data can be very good on average and still be unacceptable, because averages are the wrong thing to measure when a single collision ends the episode. Most safe reinforcement learning methods bound expected cost, which permits exactly that failure. The usual fix is to try things out in the world and learn from what goes wrong, but that is not available when the failures are the ones you were trying to avoid. So the question becomes: given a fixed dataset, no further interaction, and often no model of how the system moves, what can you actually certify? I will start from the control-theoretic answer, which is to build a function over the state space that tells you which states are still recoverable, and then keep the system inside that region. The talk covers two ways of learning such a function from static data. The first keeps it separate from the controller. V-OCBF learns a barrier value function from demonstrations and uses it as a filter that sits in front of any nominal policy and corrects actions only when needed. The second folds it into the policy itself. SafeFQL learns a safety critic that gates policy training directly, which removes the need for a dynamics model and makes deployment a single forward pass. Both are evaluated on the same navigation and locomotion benchmarks, so the comparison is concrete. I will use it to draw out which sources of error actually matter in practice, what each design costs at deployment, and where the empirical safety numbers still run ahead of anything we can prove.
Speaker Bio: Manan Tayal is a Research Scientist at Emergence AI. He completed his PhD at the Centre for Cyber Physical Systems, IISc Bangalore, advised by Shishir Kolathaya and Pushpak Jagtap, where he was a Prime Minister's Research Fellow. His thesis, Towards Trustworthy Autonomy in Uncertain Environments, develops methods for certifying the safety of learned controllers, spanning control barrier functions, Hamilton-Jacobi reachability, formally verified neural certificates, and safe offline reinforcement learning. He was most recently a visiting researcher at Microsoft Research India, and before that a Student Researcher at Stanford with Somil Bansal. He continues to collaborate with the xLab at UPenn. He did his undergraduate studies at IIT Bombay.
Youtube Recording: TBA
September 2026
September 2026
September 3: Dynamics and control of ultra-compliant mechanism systems for micro fabrication and robotics
September 3: Dynamics and control of ultra-compliant mechanism systems for micro fabrication and robotics
Speaker: Prof. Prasanna Gandhi (IIT Bombay, India)
Speaker: Prof. Prasanna Gandhi (IIT Bombay, India)
Abstract: Compliant mechanisms demonstrate numerous advantages in robotics and automation systems, including soft interaction, reduced power consumption, high durability, zero backlash, frictionless motion, and maintenance-free silent operation. However, these mechanisms present significant challenges in motion control because of complex dynamics. This work addresses these limitations through the development of compliant mechanism platforms achieving ultra-high precision positioning with high-speed capability across large motion ranges. We present a comprehensive modeling approach incorporating large deformation analysis for both static and dynamic states. We further demonstrate control of unstable equilibrium in a buckled beam-on-cart system using energy shaping techniques. The effectiveness of our approach is validated through experimental implementation, with practical applications demonstrated in SLA-based 3D microprinting and a leech-inspired robotic system.
Abstract: Compliant mechanisms demonstrate numerous advantages in robotics and automation systems, including soft interaction, reduced power consumption, high durability, zero backlash, frictionless motion, and maintenance-free silent operation. However, these mechanisms present significant challenges in motion control because of complex dynamics. This work addresses these limitations through the development of compliant mechanism platforms achieving ultra-high precision positioning with high-speed capability across large motion ranges. We present a comprehensive modeling approach incorporating large deformation analysis for both static and dynamic states. We further demonstrate control of unstable equilibrium in a buckled beam-on-cart system using energy shaping techniques. The effectiveness of our approach is validated through experimental implementation, with practical applications demonstrated in SLA-based 3D microprinting and a leech-inspired robotic system.
Speaker Bio: Dr Prasanna Gandhi is currently professor and Head of mechanical engineering, and professor-in-charge of IITB-Monash research academy and also of Suman Mashruwala Advanced Microengineering Laboratory at IIT Bombay. He received his Master’s degree from IIT Bombay (1996) and PhD from Rice University (2001), both in Mechanical engineering. He started his independent career at IIT Bombay in 2001. Current focus of his research group is on the dynamics and control of ultra-flexible mechanisms, robotics, polymer and ceramic 3D micro-printing, control of fluid instabilities for Spontaneous Multiscale Manufacturing (SMM) leading to applications in micro-printing, micro-fluidics, medical robotics, products, and devices. Some of the technologies developed in his laboratory are transferred and licensed to ISRO, a few private companies and a startup Flexmotion Technologies he cofounded recently. Among other honors, he is a recipient of Robert Lowry Patten Award at Rice University (2000), BOYSCAST fellowship of Govt of India (2006), Prof J.R.Issac fellowship (2007), Best faculty award (2008), FIE foundation award at IMTEX (2019), P K Patwardhan Technology Development Award (2021), SERB Technology Translation Award (2022), Dr Abdul Kalam Technology Innovation National Fellowship (2023), Chandrayan 3 honour (2024). He is certified teacher of stress relieving Art of Living techniques since 2010 making positive difference to lives of students, householders, and professionals alike.
Youtube Recording: TBA.
Page updated
Google Sites
Report abuse