RoboTTT Scales Robot Policy Context to 8K Timesteps
July 15, 2026
RoboTTT implements Test-Time-Training to scale visuomotor context to 8K timesteps without increasing inference latency. This enables one-shot imitation from human videos and improved performance on long-horizon, multi-stage robotic tasks.
HOW THIS AFFECTS YOU
●
builderYou can implement long-horizon robot behaviors using significantly larger context windows.
●
researcherThis demonstrates that steady gains in closed-loop performance are possible through context scaling.