Explore cutting-edge research on flexible behavior synthesis in this 40-minute presentation by MIT EECS PhD student Yilun Du. Delve into the innovative approach of planning with diffusion for generating adaptive trajectories and behaviors. Learn about neural networks combined with trajectory optimization, generative modeling for planning, and compositional trajectory generation. Discover the Diffuser model's capabilities in variable-length predictions and flexible behavior synthesis through distribution composition. Examine advanced techniques such as goal planning through inpainting, test-time cost specification, and offline reinforcement learning with value guidance. Gain insights into the application of test-time cost functions and their impact on behavior synthesis.
Overview
Syllabus
Intro
Neural nets + trajectory optimization
Is the model the bottleneck?
Planning as generative modeling
A generative model of trajectories
Compositional trajectory generation
Sampling from Diffuser
Variable-length predictions
Flexible Behavior Synthesis through Composing Distributions
Goal Planning through Inpainting
Test-Time Cost Specification
Offline Reinforcement Learning through Value Guidance
Test-Time Cost Functions
Taught by
Generative Memory Lab
