Safe Learning in Robotics

Explore the Richard M. Karp Distinguished Lecture on safe learning in robotics, delivered by Claire Tomlin from UC Berkeley. Delve into crucial topics such as air traffic control, UAV applications, and operations in unstructured environments. Discover reachable set propagation, collision avoidance techniques, and strategies for overcoming the curse of dimensionality. Learn about fast and safe planning methods, including precomputed tracking bounds and RRT applications. Examine the application of these concepts to noisily rational human models, Bayesian model confidence, and multi-agent scenarios. Investigate the use of machine learning for computing sets, safe policy gradient reinforcement learning, and online disturbance model validation. Gain insights into cutting-edge approaches for maintaining safety while learning in robotic systems.

Intro
Air Traffic Control: Separation Assurance
Growing numbers of UAV applications
Operations in unstructured environments: perceptic
Outline
Reachable Set Propagation
Level set interpretation
Numerical computation of reachable sets
Collision Avoidance Pilots instructed to attempt to collide vehicles
Backwards Reachable Set: Capture
Mode sequencing and reach-avoid
Dealing with the curse of dimensionality
Fast and Safe Planning
Precomputed Tracking Bound
10D Tracking 3D using RRT
Fast and Fast(er) Planning
Meta-Planning using FasTrack in AR/VR
Applied to: A Noisily Rational Human Model
Rationality is actually model confidence
Bayesian Model Confidence
Extending to multiple humans and robots
Analyzing introspective predictors
Using ML to compute sets
Safe Policy Gradient Reinforcement Lean
Online Disturbance Model Validation
Leaming, while staying safe