Researchers from NVIDIA's GEAR Lab, Stanford University, and the University of Texas at Austin have introduced RoboTTT (Test-Time-Training Robot Policies), a method that scales visuomotor context in robot foundation models to 8,000 timesteps — roughly three orders of magnitude beyond current state-of-the-art policies — without any increase in inference latency.
The work establishes context length as a new scaling axis for robot foundation models, analogous to how context window scaling drove breakthroughs in large language models. Findings were presented at SIGGRAPH 2026.
The core insight. At the heart of RoboTTT is the integration of test-time training (TTT) into vision-language-action (VLA) policies. Instead of storing history as a growing sequence of tokens in a Transformer, RoboTTT compresses historical experience into "fast weights" — a small set of parameters updated continuously via gradient descent during both training and inference.
The base model's "slow weights" remain frozen at inference time, while the fast weights act as a dynamic memory state that gets updated with each new observation-action pair. Because the fast weight size is fixed, single-step inference cost stays constant regardless of how long the robot has been operating.
Key technical design. Built on the GR00T N1.7 architecture, the team inserted TTT layers into the model's 16-layer DiT action head rather than training an entirely separate foundation model. To scale training to 8,000-step sequences, they combined sequence action forcing (independent noise levels for different action chunks) with truncated backpropagation through time (splitting trajectories for training while letting fast weights carry across segments).
The model was pretrained for 30,000 steps on 16 NVIDIA GB200 GPUs using desktop dual-arm robot data and first-person human video, then post-trained for an additional 20,000 steps on downstream tasks.
Results: 87% improvement on long-horizon tasks. On challenging real-robot manipulation tasks using a YAM dual-arm desktop robot with four RealSense D405 cameras, RoboTTT improved overall performance by 87% compared to a single-step context baseline. Critically, the model fully completed a five-minute, ten-stage assembly task — something no baseline method ever achieved.
Additional unlocked capabilities include:
- One-shot in-context imitation: RoboTTT successfully imitated tasks from single human video demonstrations 6 out of 10 times, while baseline methods failed entirely.
- On-the-fly policy improvement: Performance improved by 36% compared to the same model without TTT capability, even when deployed in real time.
- Robustness to perturbations: The model succeeded in 83% of trials under external disturbances, versus 53% for the best short-context baseline.
- Scaling law confirmed: RoboTTT trained with 8K timesteps outperformed the same model pretrained with 1K timesteps by 62%, providing the first evidence that closed-loop performance scales consistently with pretraining context length.
Why it matters. If context length is indeed a clean scaling axis for robot foundation models — as this work strongly suggests — it implies that simply training on longer trajectories could drive continued performance gains the same way scaling parameters and data have for LLMs. The fixed-latency property of the TTT approach means these gains come without the inference-time memory explosion that would come from simply expanding Transformer context windows. For embodied AI systems operating in the real world — where tasks unfold over minutes, not milliseconds — this could be a foundational architectural insight.
