Research Experience
Research Experience
Embedded Files
Research interests:
Research interests:
Crowd Simulation, Reinforcement Learning, Computer Graphics, Robotics, and Deep Learning.
Current Research Projects:
Current Research Projects:
Project: “Simulating Social Phenomena in Crowds with Position-Based Constraints”.
We simulate Micro-behaviors like: following, doorway, not-walking together, panic, etc.
ACCEPTED PAPERS, ABSTRACTS, POSTERS, AND PRESENTATIONS:
ACCEPTED PAPERS, ABSTRACTS, POSTERS, AND PRESENTATIONS:
Bilas Talukdar, and Tomer Weiss. Generalized, Dynamic Multi-agent Torso Crowds. In 2025 ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games (I3D 2025). May 2025. Jersey City, NJ, USA. Read the Paper here.
In crowd simulation algorithms, discs are commonly used due to their geometrical simplicity. However, discs overestimate the body’s projected shape on the ground plane, and do not naturally embody rotational information for each crowd agent. We propose a crowd simulation algorithm that represents crowd agents as 2D capsules. Our algorithm is based on position-based dynamics (PBD), for which we design novel constraints for capturing dynamics for capsule-shaped agents. We demonstrate our algorithm in several scenarios, showing superior performance compared to a disc-based representation on settings that range from sparse to dense.
In crowd simulation algorithms, discs are commonly used due to their geometrical simplicity. However, discs overestimate the body’s projected shape on the ground plane, and do not naturally embody rotational information for each crowd agent. We propose a crowd simulation algorithm that represents crowd agents as 2D capsules. Our algorithm is based on position-based dynamics (PBD), for which we design novel constraints for capturing dynamics for capsule-shaped agents. We demonstrate our algorithm in several scenarios, showing superior performance compared to a disc-based representation on settings that range from sparse to dense.
Below we list the key contributions of our work:
Below we list the key contributions of our work:
We represent each crowd agent with a torso, which realistically captures a pedestrian’s rotation and shape. This enables simulating real-world dynamic details that are difficult to capture with discs. Our approach works in settings ranging from sparse to dense.
Our method makes no assumption about the environment or specific agent active-passive behavior configurations.
We present novel torso agent-agent collision avoidance constraints by extending the position-based crowds formulation.
Our method generates dynamic crowds with varied movement and rotation behaviors, which respond to environment spacing, and mimics rotational responses observed in real-world crowds.
Bilas Talukdar, Yunhao Zhang, and Tomer Weiss. 2023. Learning Crowd Motion Dynamics with Crowds. In 2024 ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games (i3D 2024). PHILADELPHIA, PA, USA. DOI: here.
Brief Description:
Brief Description:
Currently, I am, working on a project related to improving the quality of navigation of multi-agent virtual crowds using deep reinforcement learning. In multi-agent environments, the current decision of any agent would affect the state of all other agents that are nearby it or even the future agents too. For that reason, ensuring both the collision-free and smooth virtual agent's navigation is really a challenging task. We can do that complex task by using an agent-based reinforcement learning approach. In our approach, we use an improved reward function than some existing works in the literature to ensure the smoother navigation of the virtual agents. I have trained the agents in such a way that they can exhibit the notion of real human in different complex crowd simulation scenarios (eg., Hallway, Crossway, Corridor, Circle, Obstacle etc.). My work has direct application in the field of gaming, animations, robotics, and emergency situations (eg., when fireman, police deal with emergency calls in complex scenarios). Below we list the key contributions of our work:
Currently, I am, working on a project related to improving the quality of navigation of multi-agent virtual crowds using deep reinforcement learning. In multi-agent environments, the current decision of any agent would affect the state of all other agents that are nearby it or even the future agents too. For that reason, ensuring both the collision-free and smooth virtual agent's navigation is really a challenging task. We can do that complex task by using an agent-based reinforcement learning approach. In our approach, we use an improved reward function than some existing works in the literature to ensure the smoother navigation of the virtual agents. I have trained the agents in such a way that they can exhibit the notion of real human in different complex crowd simulation scenarios (eg., Hallway, Crossway, Corridor, Circle, Obstacle etc.). My work has direct application in the field of gaming, animations, robotics, and emergency situations (eg., when fireman, police deal with emergency calls in complex scenarios). Below we list the key contributions of our work:
We design a crowd dynamics framework combining RL and position-based dynamics which propel agents to move in human-inspired fashion, observing agent acceleration and spacing.
In addition, our method consists of multiple control parameters (e.g., policy reward weights). Tuning such parameters by hand for a multi-agent simulation is difficult, since what users perceive as realistic navigational behavior may vary. Therefore, we propose a crowd-sourced, Bayesian framework to find the optimal parameters, hence the "best" policy.
Bilas Talukdar, Yunhao Zhang, and Tomer Weiss. 2024. Position-based Torso Crowds. In 2024 ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games Posters (I3D 2024 Posters). Philadelphia, PA, USA.
Bilas Talukdar, Yunhao Zhang, and Tomer Weiss. 2023. Learning to Simulate Crowds with Crowds. In ACM SIGGRAPH 2023 Posters (SIGGRAPH ’23). Association for Computing Machinery, New York, NY, USA, Article 6, 1–2.
Please find the 2-page abstract, supplementary video, and poster here.
Fillers T., Maddox, J., Ologbonyo, J., and Talukdar, B. SMC 2023 Data Challenge: EAGLE-I Explorers Solution for Challenge 7. Smoky Mountains Computational Sciences and Engineering Conference (SMC2023). Read the Paper here.
Details: In this work, we investigated five different research questions regarding power outage data across the United States at the county level. This data was collected through the Environment for the Analysis of Geo-Located Energy Information (EAGLE-I) platform at Oak Ridge National Laboratory (ORNL).
Tomer Weiss, Bilas Talukdar. A Deep Reinforcement Learning Approach Towards Human-Like Multi-Agent Dynamics. Abstract presented at: 2022 SIAM Conference on Mathematics of Data Science; September 2022; San Diego, California, U.S.
Title: A Deep Reinforcement Learning Approach Towards Human-Like Multi-Agent Dynamics.
Abstract:
Abstract:
Abstract. Controlling the motion of virtual agents with Reinforcement Learning (RL) is an area of continuing interest in multiple research communities. In this work, we focus on simulating the dynamics of multiple virtually embodied agents moving in a crowd-like manner. Crowd dynamics are important in multiple domains, such as computer vision, robotics and entertainment. When virtually capturing such dynamics, most existing RL work is concerned with directing agents towards locomotion goals in a robot-like motion. While such approaches may be effective for imitating crowds for potential robotic and entertainment applications, the resulting dynamics do not resemble human-like crowds. In this work, we propose a novel RL model which rewards agents that move in a human-like manner. Our approach is based on a novel reward function, which includes components that reward movement towards goals, a penalty for collision with other agents, and a novel human-like navigation reward, which penalizes the agent's acceleration from increasing beyond a bio-mechanically allowed threshold. We demonstrate our method's robustness on multiple crowd locomotion scenarios with varying random conditions in terms of environment, number of crowd agents, and navigation complexity, and show that our method generalizes well from one learned policy.
Abstract. Controlling the motion of virtual agents with Reinforcement Learning (RL) is an area of continuing interest in multiple research communities. In this work, we focus on simulating the dynamics of multiple virtually embodied agents moving in a crowd-like manner. Crowd dynamics are important in multiple domains, such as computer vision, robotics and entertainment. When virtually capturing such dynamics, most existing RL work is concerned with directing agents towards locomotion goals in a robot-like motion. While such approaches may be effective for imitating crowds for potential robotic and entertainment applications, the resulting dynamics do not resemble human-like crowds. In this work, we propose a novel RL model which rewards agents that move in a human-like manner. Our approach is based on a novel reward function, which includes components that reward movement towards goals, a penalty for collision with other agents, and a novel human-like navigation reward, which penalizes the agent's acceleration from increasing beyond a bio-mechanically allowed threshold. We demonstrate our method's robustness on multiple crowd locomotion scenarios with varying random conditions in terms of environment, number of crowd agents, and navigation complexity, and show that our method generalizes well from one learned policy.
Bilas Talukdar, Prasanna Balaprakash, Xingfu wu. Scaling Machine-Learning based Automatic Performance Tuning. Poster presented at: 2023 Exascale Computing Project Annual Meeting; January 2023; San Diego, Houston, Texas, U.S.
Undergraduate Thesis:
Undergraduate Thesis:
Title: An Efficient Technique to Select Features Using Naïve Bayes Method Based on Differential Evolution Algorithm and Comparison with Other Techniques.
Abstract:
Classification is an important branch of data mining. Naïve Bayes classifier is a broadly used classifier because of its accuracy, efficient calculation, and theoretical foundation. But the independent assumption of its attributes limits its performance. This study presents an approach named Differential Evolution Naïve Bayes classifier (DE-NB) which combines both Naïve Bayesian classifier and Differential Evolution algorithm to take advantage of improving classification accuracy of Naïve Bayes classifier by using an appropriate feature selection approach. This approach first searches out for an optimal subset of features using the Differential Evolution algorithm and later constructs a Naïve Bayes classifier on the reduced subset of attributes from the original attribute set. We have aimed at improving feature selection to improve the performance of the Naïve Bayesian classifier. Applying the developed DE-NB approach on seven experimental datasets, we see that in most of the cases the DE-NB outperforms the Naïve Bayes (NB), K-Nearest Neighbor (KNN), and Support Vector Machine (SVM) classifiers applied on the same Datasets.
Thesis of Students Supervised by Me:
Thesis of Students Supervised by Me:
Title: An Automated Object Detection and Emergency Case Selection (Fire) Based Street Light Platform Using Solar Power.
Abstract:
Nowadays traditionally high-intensity discharge lamps (HID) used for an urban street light on the highways remain the whole night. Energy wastes a lot when there is no vehicle movement on road. This project is improved to detect the vehicle movement on the highways in order to switch on a block of street lights ahead of the vehicle moving towards them as well as switch off the lights when the vehicle passes away using IR sensor technology to save energy. In this project, we use a microcontroller named Arduino pro mini. Using IR solar panel we will charge a 8V battery. We use a microcontroller named Adriano pro mini. Using a 12v solar panel we will charge 8Vbattery. We convert this voltage to 5v by using a voltage regulator because the system of this microcontroller will operate in 5V. The IR sensors are placed on either side of the road lamp post to sense the vehicle movement and to send the logic commands to the microcontroller to switch on/off a set of LEDs. The project operates in three modes: complete ON mode, medium mode, and dim mode. This system changes the intensity of the lights are to save lots of energy. We also add the thermal sensor in this project for fire detection so that any nearby fire can be detected by thermal sensor through the alarm. With the help of a thermal sensor we can stop or we can take immediate action in a dangerous situation in a short time
Page updated
Google Sites
Report abuse