ABSTRACT

Guaranteeing safety in autonomous systems is of paramount importance. However, providing such guarantees under various system uncertainties comes with its own challenges. In this talk, I'll describe two safe control design frameworks, where safety is modeled using Control Barrier Functions (CBF). In the first control-flavored work, we provide a risk-tunable safe control design for nonlinear systems under environmental uncertainties, for applications where a small risk in uncertainty is often tolerable. In the second constrained reinforcement learning (RL) work, we provide a provably safe RL method for nonlinear dynamical systems. This is a first of its kind result which guarantees both safety during exploration and convergence in the policy updates. Both these approaches were validated on our own custom-made safe quadrotor gym.

BIOGRAPHY

Vipul is a 3rd year Purdue IE OR PhD student with Professor Sivaranjani Seetharaman, with a focus on policy gradient and safe learning-based control design. Before Purdue, he was an optimization engineer at Siemens India, working on electricity markets. He holds master’s and bachelor’s degrees, both in electrical engineering, from the Indian Institute of Science and MJPRU, respectively. He also holds interest in cognitive sciences and hopes to combine it with mathematical tools to study behavioral dynamics in social contexts as part of his long-term research.