RoboTTT: Context Scaling for Robot Policies

Source

Source name: arXiv
Source kind: paper
Authors / organization: Yunfan Jiang, Yevgen Chebotar, Ruijie Zheng, Fengyuan Hu, Yunhao Ge, Jimmy Wu, Tianyuan Dai, Scott Reed, Li Fei-Fei, Yuke Zhu, Linxi 'Jim' Fan
Venue: arXiv: cs.RO
Published: 2026-07-16
Collected: 2026-07-18
URL: http://arxiv.org/abs/2607.15275v1
PDF: https://arxiv.org/pdf/2607.15275v1
DOI: UNKNOWN

Plain-Language Brief

The research introduces RoboTTT, a robot model and training method that enables visuomotor context scaling to 8K timesteps, far beyond current state-of-the-art policies. This allows new robot capabilities like one-shot imitation from human video demonstrations, on-the-fly policy improvement, robustness to perturbations, and stronger performance on multi-stage tasks. The method integrates Test-Time Training into robot foundation models, improves overall performance by 87% on real-robot manipulation tasks, and suggests that context length could be a new scaling axis for robot foundation models.

What the Document Says

The paper presents RoboTTT, a robot model and training method that allows visuomotor context scaling to 8K timesteps.
With this long context, the authors claim new capabilities like one-shot imitation, on-the-fly policy improvement, robustness, and stronger performance on multi-stage tasks.
Their method integrates Test-Time Training into robot foundation models, resulting in a sequence model with fast weights that compress histories and retrieve contextual information for long-context conditioning.
To scale training context length, they use sequence action forcing and truncated backpropagation through time.
On real-robot manipulation tasks, RoboTTT improves overall performance by 87% over the single-step context baseline and completes a five-minute, ten-stage assembly task which no baseline ever did.
They also find that RoboTTT trained with 8K-timestep context outperforms the same model pretrained with 1K timesteps by 62%.

Key Concepts Mentioned

RoboTTT
visuomotor context
Test-Time Training
robot foundation models
Vision-Language-Action policies
sequence model
fast weights
single-step context
real-robot manipulation tasks

Claims Made by the Document

RoboTTT scales visuomotor context to 8K timesteps without increasing inference latency.
With this scale, new capabilities are unlocked such as one-shot imitation, on-the-fly policy improvement, robustness, and stronger performance on multi-stage tasks.
RoboTTT improves overall performance by 87% over the single-step context baseline on real-robot manipulation tasks.
RoboTTT trained with 8K-timestep context outperforms the same model pretrained with 1K timesteps by 62%.

Evidence Used

The paper provides details about the RoboTTT model and training method.
Results on real-robot manipulation tasks are presented to demonstrate the performance improvement.

Methods / Approach

Test-Time Training
sequence action forcing
truncated backpropagation through time

Limitations Stated by the Authors

UNKNOWN (The paper does not discuss potential limitations of their method explicitly.)

Limitations Noted by the Agent

Agent interpretation — not from the source.

Potential limitations could include the computational resources required for training and inference at such a large context length, as well as the generalization ability of the model to various tasks and environments.

Open Questions

How does the RoboTTT method perform on other types of tasks beyond real-robot manipulation?
What are the long-term implications of scaling visuomotor context in robot policies?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The paper focuses on real-robot manipulation tasks and does not provide comprehensive performance results for other types of tasks. The potential limitations discussed here are based on common issues found in similar deep learning approaches.

Raw Abstract / Excerpt

Recent robot foundation models operate with single-step or short-history visuomotor context. We introduce Test-Time-Training Robot Policies (RoboTTT), a robot model and training recipe that scale visuomotor context to 8K timesteps, three orders of magnitude beyond state-of-the-art policies, without growing inference latency. At this context length, we unlock new robot capabilities: one-shot in-context imitation from human video demonstrations, on-the-fly policy improvement, robustness to perturbations, and stronger performance on multi-stage, long-horizon tasks. We also observe, for the first time, steady gains in closed-loop performance as pretraining context length scales. At its core, RoboTTT integrates Test-Time Training into robot foundation models such as Vision-Language-Action policies, yielding a sequence model whose recurrent state consists of fast weights, parameters updated by gradient descent during both training and inference, compressing histories into weight space and retrieving contextual information for long-context conditioning. To scale training context length, the recipe combines sequence action forcing with truncated backpropagation through time. On challenging real-robot manipulation tasks, RoboTTT improves overall performance by 87% over the single-step context baseline and fully completes a five-minute, ten-stage assembly task, which no baseline ever does. RoboTTT trained with 8K-timestep context outperforms the same model pretrained with 1K timesteps by 62%, suggesting context length as a new scaling axis for robot foundation models. Videos are available at https://research.nvidia.com/labs/gear/robottt/