Reproduce “Domain Adaptation of Visual Policies with a Single Demonstration” on OpenResearch

About this paper

Weiyao Wang, Gregory D. Hager

PromptAdapt, a Transformer-based framework, enables visuomotor policies to adapt to visual domain shifts in target environments by conditioning on a single expert demonstration. This approach allows rapid, in-context adaptation without explicit policy weight fine-tuning, leading to improved performance in both simulated and real-world robotic tasks.

Key insights

• A single expert demonstration can effectively serve as a prompt for in-context visual domain adaptation, allowing policies to adapt to new environments without requiring weight fine-tuning.
• The Transformer architecture, when conditioned on both visual observations and actions within a demonstration, can learn to generalize and adapt to complex visual changes (e.g., lighting, texture, camera pose).
• The quality and content of the demonstration are crucial; expert observation-action sequences are essential for successful adaptation, outperforming student's own rollouts or random demonstrations.

Summary by alphaXiv.