Manish Saroya

Roadmap Learning for Probabilistic Occupancy Maps with Topology-Informed Growing Neural Gas (RA-L and ICRA 2021)

Developed topology-informed Growing Neural Gas to generate navigation roadmaps for uncertain environments. Compared to a denser PRM, our method provides a ×5 computation efficiency improvement for shortest path search.

Figure shows a roadmap (grey graph) generated by our topology-informed growing neural gas algorithm. The probabilistic occupancy grid (colored image) is generated using a Hilbert map from point-cloud observations of the Intel Research Lab. The white path is a solution path planned over the roadmap through uncertain obstacles and narrow doorways.

Online Exploration of Tunnel Networks Leveraging Topological CNN-based World Predictions

IROS 2020 Paper, Code, Presentation Video

Learning structural cues to gain insight about unexplored regions in subterranean tunnel networks using techniques from topological image segmentation and image inpainting. These topological predictions improve frontier based exploration policy by 11–30 %.

In the figure, robot (blue) navigates from the start (red) at bottom centre. At each timestep, the robot decides which frontier (orange) to navigate to next. For the top row, the areas predicted to be open are shown in green. These lie within the unmasked area of interest (shaded).

Utilizing Temporal Information for Underwater 3D Reconstruction from Sonar Images

Paper
Developed a novel recurrent convolutional deep learning architecture forestimating the missing elevation angle from 2D sonar images. Recurrent approach improves performance by up to8 % over previous single-frame deep learning approaches.

Responsible for Localization Stack | The Hitech Robotics Systemz Ltd.

Video,

Integrated 2D lidar-based Adaptive Monte Carlo Localization (AMCL) on Autonomous Ground Vehicles (AGVs) utilized for material handling applications in various factory floors. These AGVs are in continuous operation since deployment in 2018.

Controller for Autonomous Forklifts | The Hitech Robotics Systemz Ltd.

Video,

Implemented and deployed a pure pursuit path tracking algorithm on autonomous forklifts. Generated non-holonomic trajectories for forklifts to perform pallet picking tasks with the path-planning team.

Research Patent - Optimized Odometry Motion Model Parameters for Forklifts |
The Hitech Robotics Systemz Ltd.

Devised a method for learning odometry motion model parameters to account for manufacturing errors.

Generated Tunnel and Urban Subterranean Environments

Programmatically generated realistic subterranean tunnel networks. These environments contain structure in the form of topological features (loops and connected passages).

The below figure shows programmatically generated 3D multi-level gazebo world that are intended to represent the SubT urban circuit environment. Each generated world consists of a set of randomized floors. Each floor contains a random number of walls, with random sizes, positions and visual textures. Each floor plane has rectangular holes cut out of it so that a robot can navigate between different levels.

Multiagent Controller | CMU | Advisor: Prof. Howie Choset

Website

Designed a multiagent sequential controller utilizing a smooth affine field for robustness against environmental perturbations and modeling errors.

Motion Planning for Omni-directional Spherical Modular Snake Robot (OSMOS)

Paper

Developed a snake robot by mechanically connecting multiple spherical modules. Derived kinematic model for state estimation and control.

Snake Robot -- Design and Control

Report

3D printed 9 DOF Snake Robot, which could navigate in rocky terrains. Wrote ROS nodes to perform various gaits such as sidewinding, rolling, and tree climbing. Created a GUI with Python's TkInter to run different gaits, and control various parameters such as motor-ID, speed, phase, etc.

Control of `Swayat'-20 DOF Humanoid (Kid Size)

Generated joint angles for walking and grasping gaits.