Baiyu Peng

Research Interests Summary

Constraint Learning

Robot tasks are governed by constraints such as obstacle clearance, tool orientation, contact geometry, and stage-specific requirements. These constraints can be difficult to specify by hand. My doctoral research learns explicit constraint models from demonstrations and uses them for motion planning and control, so that robots can adapt to new goals and task conditions while respecting the learned requirements.

I study how to learn from feasible and failed demonstrations, select informative expert queries, recover equality constraints, and identify changing constraints in multi-stage tasks.

• Positive-Unlabeled Constraint Learning (PUCL): inferring nonlinear continuous constraints from feasible demonstrations.
• Actively Learning Task Constraints: combining feasible and failed demonstrations with active expert queries.
• Learning Equality Constraints from Trajectory Demonstrations (ONCL): learning constraint manifolds for planning and manipulation. [video]
• Stage-Wise Constraint Learning (SWCL): discovering stage boundaries and stage-specific constraints from unsegmented demonstrations. [video]

Safe Reinforcement Learning

Reinforcement Learning (RL) has shown exceptional successes in a variety of domains like Go, video Games and robotics, but its application in real world is still limited due to lack of safety guarantee. I am interested in developing safe RL algorithms with explicit safety constraints, promoting the RL in the real world rather than only simulators and games.

Obstacle-avoidance via safe RL. See full video here.

Autonomous Driving

Autonomous driving vehicles are expected to have a huge positive impact on the automotive industry. I am especially interested in applying the RL algorithms in decision-making and control of the autonomous vehicles and mobile robot.

End-to-end autonomous driving via RL. See full video here.

Multi-Robot Distributed Planning and Control

The use of autonomous mobile robots is capable of replacing human labor for demanding, repetitive tasks. The key challenge is the control of multiple robots interreacting with each other. I am interested in applying the RL and model predictive control method to enable multi-robot distributed planning and control.

Multi-robot control via distributed MPC. See full video here.