RoboTTT scales robot policy context to 8K timesteps — three orders of magnitude beyond state of the art — via test-time training

arXiv·medium signal

Submitted July 16 by a Stanford/NVIDIA-affiliated team including Li Fei-Fei, Yuke Zhu, and Jim Fan, RoboTTT uses fast weights updated by gradient descent at inference to compress visuomotor history, combined with sequence action forcing and truncated backprop through time. It delivers 87% improvement over a single-step baseline on real-robot manipulation and completes a five-minute, ten-stage assembly task no baseline finished. Models with 8K-timestep context beat 1K-context models by 62%, and the approach unlocks one-shot imitation from human video.