NVIDIA's RoboTTT stretches robot context to 8K timesteps

For a couple of years the interesting axis for robot foundation models has been the same one everyone argues about for language models: more data, more parameters. A new paper from NVIDIA, Stanford and UT Austin, called RoboTTT, argues that context length deserves a seat at that table, and puts a real-arm number on it.

The system stretches visuomotor context to 8K timesteps, which the authors describe as three orders of magnitude beyond state-of-the-art robot policies. On the paper's own benchmarks, RoboTTT reportedly improves overall performance by 87% over a single-step context baseline, and the version trained with 8K-timestep context beats the same model pretrained with 1K timesteps by 62%. The headline demo, according to the project page, is a five-minute, ten-stage assembly task that no baseline ever completes.

The mechanism worth understanding is Test-Time Training. Rather than freezing weights at inference, RoboTTT's recurrent state is made of fast weights that keep updating as the robot runs, and the model is trained by combining sequence action forcing with truncated backpropagation through time. The authors' claim is that this scales context without growing inference latency, which is where naive long-context transformers tend to fall over on a physical arm.

Why this matters if you are not building manipulation policies yourself: the tasks in the paper (Pup Go Car, Gear Bot, Circuit) are the kind of long-horizon assembly work that has been the wall for robot policies. Longer context is what lets a policy remember what it already did in step three when it gets to step nine, and it is also what the authors say unlocks the more speculative capabilities they demonstrate, one-shot in-context imitation from a human video and on-the-fly policy improvement after a perturbation.

The honest caveat is that this is a single paper on its own real-arm benchmarks, from the group that most benefits from a new scaling axis being taken seriously. The reporting doesn't tell you how the recipe transfers to hardware the team did not tune for, or how fast-weight state generalises past these three tasks, or what the training bill actually looks like. But if context length really is a third scaling axis for physical AI, alongside model size and data, that changes the conversation about what a robot foundation model even is.