NVIDIA researcher Jim Fan unveils RoboTTT robot policy with 8,000-timestep context and test-time training

NVIDIA researcher Jim Fan posted on X on July 15 that his team scaled a robot model to 8,000 timesteps of context with constant inference cost. Fan said the new policy, called RoboTTT, uses test-time training to carry a tiny model inside the main model. He said every incoming sensor reading triggers one gradient step on that tiny core so the history keeps getting compressed into its weights. Fan said the hidden state has a fixed size so the robot can handle arbitrarily long experience with little overhead and learning continues indefinitely after deployment. He said RoboTTT enables one-shot in-context learning from human video and showed self-improvement on the fly with the robot recovering from its own errors mid-episode. Fan said a context scaling curve from 128 to 8,000 timesteps showed closed-loop performance hill-climbing steadily with no sign of saturation and that 8K-context pretraining beats 1K by 62%.