Projects
Current Projects
RAIL: Reachability-Aided Imitation Learning for Safe Policy Execution
Wonsuhk Jung, Dennis Anthony, Utkarsh Mishra, Nadun Ranawaka, Matthew Bronars, Danfei Xu *, Shreyas Kousik *
Guaranteed Reach-Avoid for Black-Box Systems through Narrow Gaps via Neural Network Reachability
Long Kiu Chung, Wonsuhk Jung, Srivatsank Pullabhotla, Parth Shinde, Yadu Sunil, Saihari Kota, Luis Felipe Wolf Batista, Cédric Pradalier, Shreyas Kousik
Chuizheng Kong, Alex Qiu *, Idris Wibowo *, Marvin Ren, Aishik Dhori, Kai-Shu Ling, Ai-Ping Hu, Shreyas Kousik
Goal-Reaching Trajectory Design Near Danger with Piecewise Affine Reach-avoid Computation
Long Kiu Chung *, Wonsuhk Jung *, Chuizheng Kong, Shreyas Kousik
Robotics: Science and Systems (RSS), 2024
Current Research Directions
Our overall goal is to ensure full-stack safety by studying each component of the autonomy stack. We seek to build shared representations of uncertainty for each component such that, in the long term, a robot can teach itself to be safe.
Past Work
Reachability-Based Trajectory Design
Reachability-Based Trajectory Design, or RTD, is a receding-horizon planning method that generates dynamically-feasible, collision-free trajectories for autonomous mobile robots. Check out the tutorial for a walkthrough.
Modelling Uncertainty in Estimators and Learned Models
We developed a computationally-efficient method for shadow-matching, where one uses a 3-D urban map to identify GNSS (Global Navigation Satellite System) shadows, or areas where satellite signals are blocked, to create artificial set-valued measurements that represent uncertain possible receiver positions. To represent the curved convex shapes in this method, such as the Gaussian distribution confidence ellipsoids commonly associated with measurement uncertainty, we also created ellipsotopes, a novel set representation that fuses the benefits of polytopes and ellipsoids.
To compute reachability for learned models, we used RTD to build a safety layer for a reinforcement learning (RL) agent, which can outperform vanilla RTD. We also computed the exact forward reachable sets of feedforward neural networks, bridging the gap between verification and training of a neural network.