SciGroveBeta
Robotics

RoboTTT: Context Scaling for Robot Policies

Yunfan Jiang, Yevgen Chebotar, Ruijie Zheng, Fengyuan Hu, Yunhao Ge, Jimmy Wu, Tianyuan Dai, Scott Reed, Li Fei-Fei, Yuke Zhu, Linxi Fan

Featured July 21, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

This paper introduces a robot brain, RoboTTT, that can remember and learn from very long past experiences, like five minutes of video, by constantly updating special "fast weights" on the fly, making robots much smarter and more adaptable.

In depth
The paper introduces RoboTTT, a novel robot policy that significantly scales visuomotor context to 8K timesteps by integrating Test-Time Training (TTT) into robot foundation models. This allows the policy's recurrent state to consist of fast weights, which are dynamically updated via gradient descent during both training and inference, enabling the model to compress long histories and adapt on-the-fly without increasing inference latency.

Key Takeaways

  • 1
    RoboTTT scales visuomotor context to 8K timesteps, three orders of magnitude beyond prior state-of-the-art, unlocking new robot capabilities like one-shot imitation and on-the-fly improvement.
  • 2
    The core innovation is integrating Test-Time Training (TTT), where "fast weights" are updated during inference, allowing dynamic adaptation and efficient context compression.
  • 3
    The training recipe combines sequence action forcing and truncated backpropagation through time (TBPTT) to handle long sequences efficiently, demonstrating that context length is a new scaling axis for robot foundation models.

Conceptual Flow

HIGH LEVEL
1
Methodology (The 'Logic')

The robot learns from long past experiences by constantly updating a small part of its brain, called "fast weights," as it works.

Robot Sees
Robot Acts
Past Events
Update Small Brain Part
Smarter Robot Action
2
Results (The 'Impact')

This new method helps robots do complex, long tasks much better, learn from one human example, and fix mistakes by themselves.

Old Robot Brain
New Robot Brain
Much Longer Memory
Better Task Completion
Learns from One Demo
Fixes Mistakes