Long context windows are a new scaling axis for robot learning: policies trained with 8K-timestep history outperform 1K-timestep versions by 62%, unlocking capabilities like one-shot imitation and on-the-fly adaptation that short-context robots lack.
RoboTTT enables robot policies to use 8,000 timesteps of visual and action history—thousands of times longer than current systems—without slowing down inference. By combining test-time training with efficient memory compression, the system learns from human video demonstrations on-the-fly, recovers from disruptions, and completes complex multi-step tasks that baseline robots cannot finish.