Explore recent advancements in machine autonomy and generalizable embodied AI through this remote talk given at Stanford Graphics Group. Delve into topics such as sequential decision making in complex environments, procedural generation of new environments, and benchmarking reinforcement learning generalization. Learn about innovative approaches like Real2Sim for generating traffic scenarios, policy pretraining with real-world data, and self-supervised learning through contrastive methods. Discover the potential of human-in-the-loop reinforcement learning and human-AI copilot optimization. Gain insights into policy dissection through frequency analysis and witness a demo of learning to drive in the CARLA environment.
Overview
Syllabus
Intro
Sequential Decision Making in Complex Environments
Data from the Existing Simulation Environments
Procedural Generation of New Environments
Benchmarking RL Generalization
Benchmarking Safe Reinforcement Learning
Benchmarking Multi-Agent Reinforcement Learning
Real2Sim: Learning to generate traffic scenarios
Pretraining Policy Representation with Real World Data
Self-supervised Learning through Contrastive Learning
Policy Pretraining with Human Actions
Action-conditioned Contrastive Learning
Pretrained Representation for Imitation Learning
Human-in-the-loop Reinforcement Learning
Human-Al Copilot Optimization (HACO)
Demo Video: Learning to drive in CARLA environment
Policy Dissection through Frequency Analysis
Taught by
Bolei Zhou
