SURE trains robot vision-language-action models with a tenth of the demonstration data

Researchers led by Jingye Zhang have described SURE in an arXiv paper, a training framework for vision-language-action robot models that accepts that several nearby actions can solve the same task and that demonstrations unfold at different speeds. It adds a light probabilistic action decoder to capture spatial uncertainty and Soft-DTW alignment to allow for timing differences.

On LIBERO, RoboTwin 2.0 and real-world manipulation tasks, SURE matched or beat baseline models when trained on just a tenth of the demonstrations or a fifth of the usual training steps, the authors report.