RoboTTT: Robot Context Extended to 8K Steps
nvidia · hf · 2026-07-17
RoboTTT: Scaling Robot Policy Context to 8K Steps
NVIDIA proposed RoboTTT (Test-Time-Training Robot Policies), expanding the visuomotor context of robot policies to 8K timesteps—three orders of magnitude higher than existing policies, without increasing inference latency.
Method
- Integrates Test-Time Training into robotic foundation models, especially Vision-Language-Action policies
- The recurrent state of the sequence model isn't a standard hidden state, but fast weights updated via gradient descent
- Uses sequence action forcing + truncated BPTT during training to extend context length
Capabilities Unlocked
- One-shot in-context imitation from human video demonstrations
- Online policy improvement
- Increased robustness against perturbations
- Stronger long-horizon, multi-stage task capabilities
Results
- On real-world robotic manipulation tasks, overall performance improved by 87% over single-step context baselines
- Successfully completed a 5-minute, 10-step assembly task, which baselines failed to do
- Models pretrained with 8K context further improved by 62% compared to the 1K version
The authors explicitly position "context length" as a new scaling axis for robotic foundation models.
Related event: RoboTTT brings test-time training to robot policies(6 posts)→
More from Embodied
- Robotics paper says VLA and world models are not enough for grounded supervision — hbouammar · 2026-07-23
- Tesla Active FSD Subscriptions Surge 56% YoY to 1.48 Million — Polymarket · 2026-07-23
- NASA adds $1.2M to Motiv Space Systems’ robot contract — BrettKrieger12 · 2026-07-23
- SIGGRAPH 2026 method turns 3D objects into multi-touch sensors — cholz · 2026-07-23
- Patch Policy preserves dense spatial detail for robot manipulation — chris_j_paxton · 2026-07-23
- Robot “game skins” become real-world wearable armor at REK Tokyo — cixliv · 2026-07-23