Stanford And Nvidia Researchers Introduce RoboTTT Robot Model Extending Context Memory To 8,000 Timesteps

Stanford and Nvidia researchers introduce RoboTTT, a robotic learning framework that extends a system's operational context to 8,000 timesteps, or 5 minutes of continuous memory, while keeping inference costs constant. The architecture uses a test-time training method that embeds a miniature neural net inside the primary policy, applying a gradient step for every sensor reading to compress history directly into the model's weights.

The method enables one-shot imitation from human demonstration videos and allows the system to self-correct mid-episode by distilling error recovery into its context. The researchers report that scaling the context from 128 to 8,000 timesteps lifts closed-loop performance by 62% compared to 1,000-timestep baselines, showing a consistent scaling curve with no sign of saturation.