RoboTTT: Context Scaling for Robot Policies
RoboTTT pushes robot policy context to 8,000 timesteps without adding inference latency.
The paper says that longer visuomotor context enables one-shot imitation from human videos, on-the-fly policy improvement, perturbation robustness, and better long-horizon task performance. Its recurrent state uses fast weights updated by gradient descent during training and inference, compressing history into the model’s parameters. On real-robot manipulation tasks, RoboTTT improves overall performance by 87% over a single-step baseline. The authors report it fully completes a five-minute, ten-stage assembly task that no baseline completes, and that 8K-context training beats 1K-context training by 62%.
ArXiv · AI/CL/LG's note
The paper says that longer visuomotor context enables one-shot imitation from human videos, on-the-fly policy improvement, perturbation robustness, and better long-horizon task performance. Its recurrent state uses fast weights updated by gradient descent during training and inference, compressing history into the model’s parameters. On real-robot manipulation tasks, RoboTTT improves overall performance by 87% over a single-step baseline. The authors report it fully completes a five-minute, ten-stage assembly task that no baseline completes, and that 8K-context training beats 1K-context training by 62%.
ArXiv · AI/CL/LG's note
score 7