Researchers led by Chengyang Dong have built Dune, an optimiser that adapts vision-language-action models to new conditions while a robot is in use, according to a paper on arXiv. Backpropagation needs a lot of memory on small robot computers. Zeroth-order methods avoid it, but their noisy estimates tend to stall over long runs.
Dune learns a small predictor of errors from past optimisation steps and combines it with the current estimates. This reduces variance without backpropagation or extra passes through the model. The authors report higher success rates with a small memory footprint on the LIBERO benchmark and on real robots, with steadier control and better recovery after disturbances.



