RoboTTT: Context Scaling for Robot Policies

arXiv cs.AI · 6 hours ago

Researchers introduced RoboTTT, a robot policy model that extends visuomotor context to 8,000 timesteps, compared to single-step or short-history approaches used in current robot foundation models. The model improves performance by 87% over single-step baselines and completes a ten-stage assembly task lasting five minutes that baseline models cannot complete. This context scaling enables robot policies to learn from longer visual histories during inference, allowing capabilities like one-shot learning from human videos and improved performance on multi-stage tasks.