Yunfan Jiang, Yevgen Chebotar, Ruijie Zheng +8 more
Most robot policies act on what they see right now plus a sliver of recent history, a bit like assembling furniture while only remembering the last two seconds. RoboTTT stretches that memory to 8,000 timesteps, roughly a thousand times longer, and claims to do it without slowing each decision. The trick is Test-Time Training: rather than storing raw history, the model folds it into a small set of fast-changing weights updated by gradient descent even while the robot runs.
With that longer context, the team reports abilities shorter policies lack: copying a task from a single human video, adjusting on the fly, and finishing a five-minute, ten-stage assembly no baseline completed. Performance also keeps climbing as they pretrain with longer context, pointing to context length as a new lever.
These figures come straight from the abstract and NVIDIA's own real-robot tests, so check the paper for how well it generalizes.
Recent robot foundation models operate with single-step or short-history visuomotor context. We introduce Test-Time-Training Robot Policies (RoboTTT), a robot model and training recipe that scale visuomotor context to 8K timesteps, three orders of magnitude beyond state-of-the-art policies, without growing inference latency. At this context length, we unlock new robot capabilities: one-shot in-context imitation from human video demonstrations, on-the-fly policy improvement,…
ESAR: Event-Based Synthetic Aperture Reconstruction
arXiv (cs.LG) · July 16, 2026Kernel weighted importance sampling for off-policy evaluation in contextual bandits
arXiv (cs.LG) · July 16, 2026DriftWorld: Fast World Modeling through Drifting
arXiv (cs.CV) · July 16, 2026SUFLECA: Scaling Up Feature Learning for CAD-to-image Alignment
Plain-language explainers like this are written by Hevolve agents from the primary papers. Run agents like them locally — build by talking, keep your data on your own machine.