RoboTTT: Test-Time Training for 8K-Timestep Robot Policy Context

RoboTTT uses test-time training and fast-weight recurrence to scale robot policy context to 8K timesteps without reported growth in inference latency.

TL;DR — RoboTTT is a robot model and training recipe that uses Test-Time Training and fast weights to scale visuomotor context to 8K timesteps. The abstract reports no growth in inference latency, an 87% overall performance gain over a single-step context baseline, a 62% gain for 8K-context pretraining over 1K-context pretraining, and full completion of a five-minute, ten-stage assembly task that no baseline completes.

Core contribution

RoboTTT targets a bottleneck identified directly in the abstract: recent robot foundation models usually condition on only a single step or a short visuomotor history. That can be limiting for tasks where earlier demonstrations, stages, corrections, or disturbances matter later in execution.

The abstract's central claim is that RoboTTT scales usable visuomotor context to 8K timesteps, described as three orders of magnitude beyond state-of-the-art policies, without growing inference latency. It presents context length itself as a scaling axis for robot foundation models.

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.