NVIDIA introduces RoboTTT, a robot foundation model and training recipe that extends visuomotor context to 8,000 timesteps without increasing inference latency. By integrating Test-Time Training into Vision-Language-Action policies, the system compresses history into weight space using fast weights updated by gradient descent.

  • Enables one-shot in-context imitation from human video demonstrations and on-the-fly policy improvement.
  • Achieves robustness to perturbations and stronger performance on multi-stage, long-horizon tasks.
  • Improves overall performance by 87% over single-step context baselines on real-robot manipulation tasks.
  • Fully completes a five-minute, ten-stage assembly task that no baseline could finish.
  • Outperforms models pretrained with 1,000 timesteps by 62%, establishing context length as a new scaling axis.

The authors consider this important because it unlocks new capabilities for robot foundation models and demonstrates steady gains in closed-loop performance as pretraining context length scales.