WAM-TTT: Teaching Robots to Adapt to New Environments by Watching Humans | GALBOT AstraBrain

In this talk

1. 0:26

   Why deployment is difficult

   Embodied foundation models are first pretrained and then post-trained, often with data gathered in dedicated training centers. When a robot moves to a new environment, its performance can drop sharply. Teleoperation is costly, slow, and requires specialized equipment, while the resulting skills may not generalize well.

2. 0:56

   Adapting from human videos

   WAM-TTT adapts robots using recorded human demonstrations, without action labels or robot action data collected at deployment. In an ablation study, combining 100 robot trajectories with 100 human videos produced a 74.1% average task completion score. That result was comparable to, and better than, using robot data alone.

3. 1:37

   Keeping the foundation model intact

   The method learns from raw human video rather than relying on detailed action labels or precise motion retargeting. Its foundation model stays frozen while lightweight fast weights are updated for the new environment. This is intended to reduce catastrophic forgetting and retain the model’s existing generalization ability.

4. 2:30

   Cross-environment results

   Evaluations varied lighting, workspaces, and object instances. Using human videos only as contextual input retained 14.7% performance, while WAM-TTT retained 75.6%. The results favor fast-weight adaptation over relying solely on in-context learning.