RoboTTT Enables Long-Memory Robot Learning

DrJimFan · x · 2026-07-15

Introducing RoboTTT: applying test-time training to robotic policies. By performing gradient updates on a small kernel network during inference, it compresses historical experience into the model weights.

Key results include:

Natively extends the robot's context length to 8,000 timesteps, equivalent to about 5 minutes of "muscle memory," while keeping inference costs constant.
Unlike older policies that process a few frames at a time and quickly forget, this approach allows the robot to learn continuously and improve even after deployment.
Supports one-shot in-context learning: after a single human video demonstration, the robot can imitate previously unseen circuit board assembly configurations.
If an error occurs during execution, the robot can incorporate the "corrective action" into its context, enabling self-correction.

The authors also report a new context scaling curve: from 128 to 8K timesteps, closed-loop performance consistently improves without saturating; 8K context pre-training boosts performance by 62% compared to 1K.

Original post →

RoboTTT Enables Long-Memory Robot Learning

More from Embodied