RoboTTT: Context Scaling for Robot Policies
Merged summary
TL;DR - RoboTTT is a robot foundation model and training recipe that scales visuomotor context to 8K timesteps (3 orders of magnitude beyond prior policies) using Test-Time Training, unlocking long-horizon and in-context capabilities without added inference latency. It matters because it introduces context length as a new scaling axis for robot policies.
- Integrates Test-Time Training into Vision-Language-Action policies: a sequence model whose recurrent state is "fast weights" updated by gradient descent during both training and inference, compressing histories into weight space for long-context conditioning.
- Training recipe combines sequence action forcing with truncated backpropagation through time to reach 8K-timestep context.
- Reports 87% overall improvement over a single-step baseline on real-robot manipulation, and full completion of a five-minute, ten-stage assembly task that no baseline completes.
- 8K-timestep pretraining outperforms the same model at 1K timesteps by 62%, and unlocks one-shot in-context imitation from human video, on-the-fly policy improvement, and robustness to perturbations.
Sources (1)
RoboTTT: Context Scaling for Robot Policies
TL;DR - RoboTTT is a robot foundation model and training recipe that scales visuomotor context to 8K timesteps (3 orders of magnitude beyond prior policies) using Test-Time Training, unlocking long-horizon and in-context capabilities without added inference latency. It matters because it introduces context length as a new scaling axis for robot policies.
- Integrates Test-Time Training into Vision-Language-Action policies: a sequence model whose recurrent state is "fast weights" updated by gradient descent during both training and inference, compressing histories into weight space for long-context conditioning.
- Training recipe combines sequence action forcing with truncated backpropagation through time to reach 8K-timestep context.
- Reports 87% overall improvement over a single-step baseline on real-robot manipulation, and full completion of a five-minute, ten-stage assembly task that no baseline completes.
- 8K-timestep pretraining outperforms the same model at 1K timesteps by 62%, and unlocks one-shot in-context imitation from human video, on-the-fly policy improvement, and robustness to perturbations.