LatentCBF: A Control Barrier Function in Latent Space for Safe Control
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Representation Learning, Reinforcement Learning, Optimal Control, End-to-End Learning, Convex Optimization, Control Barrier Function, Autonomous Driving, CARLA, Robotics
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Safe control is crucial for safety-critical autonomous systems that are deployed in dynamic and uncertain environments. Quadratic-programming-control-barrier-function (QP-CBF) is becoming a popular tool for safe controller synthesis. Traditional QP-CBF relies on explicit knowledge of the system dynamics and access to all states, which are not always available in practice. We propose LatentCBF (LCBF), a control barrier function defined in the latent space, which only needs an agent's observations, not full states. The transformation from observations to latent space is established by a Lipschitz network-based AutoEncoder. In addition, the system dynamics and control barrier functions are all learned in the latent space. We demonstrate the efficiency, safety, and robustness of LCBFs in simulation for quadrotors and cars.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9335
Loading