RoboTTT

RoboTTT (Test-Time-Training Robot Policies) is a VLA foundation model and training recipe introduced by NVIDIA Research, Stanford University, and the University of Texas at Austin that scales visuomotor context (history of observations, proprioception, and actions) to 8,000 timesteps — approximately 5 minutes at 30 Hz control — three orders of magnitude beyond typical single-step baselines, while keeping inference latency constant independent of context length. RoboTTT achieves an 87% improvement over single-step baselines on long-horizon tasks. It enables one-shot in-context imitation from a single human video demonstration of an unseen configuration, and supports on-the-fly self-improvement and strategic recovery from its own errors mid-episode via DAgger Distillation. Key capabilities include completing multi-stage bimanual assembly tasks (e.g., five-minute ten-stage vehicle assembly, circuit board assembly with varying configurations) that no single-step baseline could complete. RoboTTT is the first model to demonstrate consistent closed-loop performance improvement from 8,000-timestep pretraining context.