Researchers introduce RoboTTT, a training recipe that extends visuomotor context for robot foundation models to 8,000 timesteps without increasing inference latency. The method integrates Test-Time Training into Vision-Language-Action policies by using fast weights updated via gradient descent during both training and inference.
- Context length is scaled using sequence action forcing combined with truncated backpropagation through time.
- The approach enables one-shot in-context imitation, on-the-fly policy improvement, and robustness to perturbations.
- RoboTTT improves overall performance by 87% over single-step baselines and completes a five-minute, ten-stage assembly task that other models fail to finish.
- Models trained with 8K-timestep context outperform those pretrained with 1K timesteps by 62%, establishing context length as a new scaling axis.
The authors consider this significant because it unlocks capabilities for multi-stage, long-horizon tasks and demonstrates steady gains in closed-loop performance as pretraining context scales.