Jiawei (Janna) Lin
Dynamic Safety Shielding for Reinforcement Learning in Black-box Environments
Independent Advised by Prof. Hadas Kress-Gazit Shielding, Automata Learning, Reinforcement Learning
Reinforcement Learning (RL) algorithms aim to compute policies that maximize reward without necessarily guaranteeing safety during learning or execution. This makes it challenging to use RL in cyber-physical systems such as robotics, where unsafe actions may damage the robot and/or its environment. In contrast to most previous studies that propose techniques to reduce the number of unsafe actions given some prior knowledge of the environment, we are interested in the situation where undesired behaviors can be reduced without any prior system knowledge. Hence, in this study, we aim to explore a dynamic shielding technique that safeguards RL. Specifically, we implement a variant of RPNI, a type of automata learning, which is constructed in parallel with the model-based RL to filter out potential unsafe behaviors during action execution before the agent actually experiences it. Our experiment shows that the number of undesired trials can be effectively reduced compared to unshielded RL.
Research Project Image
Continuous Execution of High-level Collaborative Tasks for Heterogeneous Robot Teams
Group Advised by Prof. Hadas Kress-Gazit Formal Synthesis, Multi-Robot Communication
Abstract to be added
Research Project Image
Preference-Informed Whole-arm Manipulation for Physical Human-Robot Interaction
Group Advised by Prof. Tapomayukh Bhattacharjee Multi-priority Nullspace Control, Reinforcement Learning
Abstract to be added
Research Project Image