Abstract:
With the growing deployment of robots across various safety-critical applications, ensuring safety has become increasingly essential to prevent accidents and enhance reliability. Addressing safety in robotic systems presents several inherent challenges. Firstly, in multi-agent settings, incorporating safety filters can adversely impact overall task performance, as agents must simultaneously ensure collision avoidance and timely task completion. To mitigate this trade-off, we propose a Co-Safe Reinforcement Learning (RL) framework wherein the ego agent utilizes the neighbouring agents’ safety model information to better inform its policy, thereby minimally influencing other agents’ trajectories, resulting in improved task performance on the defined metrics. Secondly, this thesis tackles safety assurance for soft robots, whose inherently complex dynamics pose significant difficulties for traditional model-based safety filters. To address this, we present a model-free safety filter based on Q-learning, designed to seamlessly integrate with standard RL paradigms. The practical viability and robustness of the proposed safety filter is demonstrated through simulation studies and real-world validations using a soft robotic limb actuated by shape memory alloys.

Committee:
Prof. Guanya Shi (advisor)
Prof. John M. Dolan (advisor)
Prof. Changliu Liu
Simin Liu