RoboTTT: Context Scaling for Robot Policies — A New Scaling Axis for Foundation Models

· Editorial Team estimated
robot foundation models test-time training context scaling manipulation NVIDIA

NVIDIA and Stanford researchers propose RoboTTT, scaling visuomotor context to 8K timesteps — three orders of magnitude beyond prior work — enabling one-shot in-context imitation from human video, on-the-fly policy improvement, and full completion of a five-minute, ten-stage assembly task with 87% improvement over single-step baselines.

Paper · arXiv:2607.15275

RoboTTT: Context Scaling for Robot Policies — A New Scaling Axis for Foundation Models

Background

Today’s robot foundation models operate with severely limited temporal context — typically a single timestep or a short history window. This constraint fundamentally limits their ability to perceive, reason about, and plan for long-horizon tasks such as multi-stage assembly, surgical procedures, or complex manufacturing sequences where decisions depend on events that occurred seconds or minutes earlier.

Key Innovation

A joint team from NVIDIA GEAR Lab and Stanford University introduces RoboTTT (Test-Time-Training Robot Policies), which integrates Test-Time Training into Vision-Language-Action (VLA) policies to achieve unprecedented context scaling.

Three core technical contributions:

  • 8K-timestep visuomotor context: Three orders of magnitude beyond state-of-the-art policies, achieved without increasing inference latency. The key insight is casting gradient-descent-updated fast weights as the recurrent state of a sequence model, compressing history into weight space for efficient long-context retrieval.

  • Scalable training recipe: Combining sequence action forcing with truncated backpropagation through time solves the convergence challenges of ultra-long sequence training, establishing context length itself as a viable training dimension.

  • Emergent capabilities at scale: Long context unlocks qualitatively new behaviors — one-shot in-context imitation learning from human video demonstrations, on-the-fly policy correction during execution, robustness to perturbations, and significantly stronger performance on multi-stage, long-horizon tasks.

Results

On challenging real-robot manipulation benchmarks, RoboTTT achieves an 87% improvement over single-step context baselines. Critically, it is the first model to fully complete a five-minute, ten-stage assembly task — a milestone no baseline method could reach. The 8K-context model outperforms the same architecture pretrained with only 1K context by 62%, providing strong evidence that context length represents a genuine new scaling axis for robot foundation models.

Industry Implications

RoboTTT signals a paradigm shift: from “bigger models, more data” toward “longer context, stronger memory.” For industry, this translates to:

  1. Simplified teaching pipelines: Operators record a single human demonstration video — the robot learns to imitate it in one shot, dramatically reducing programming and setup costs.
  2. Online adaptation: Robots improve their policies during execution, adapting to environmental changes without halting production.
  3. Complex assembly automation: Long-horizon multi-stage assembly becomes practically feasible — a milestone for flexible manufacturing and industrial automation.

The work is led by NVIDIA GEAR Lab with authors including Yunfan Jiang, Yevgen Chebotar, Ruijie Zheng, Fengyuan Hu, Yunhao Ge, Jimmy Wu, Tianyuan Dai, Scott Reed, Fei-Fei Li, Yuke Zhu, and Linxi “Jim” Fan.

Project page: https://research.nvidia.com/labs/gear/robottt/