Safe Autonomy has many open research challenges on the critical path to launch. This talk is an overview of recent papers published by the TRI ML team and collaborators towards building a modular yet end-to-end differentiable robotics stack. First, we will see how to leverage data, redundancy, and efficiency to improve Robustness in Perception. Second, we will address the inherent Randomness in Prediction by modeling uncertainty, latent intents, and multi-modal future distributions. Finally, we will present results in Risk-aware Planning and Control, leveraging demonstrations, Reinforcement Learning, the entropic risk measure, and causality.
Copyright. All Rights Reserved.