Derived from Cosmos Policy, SmolWAM reduces the computation and memory required. It achieves 95.5% success over 400 LIBERO episodes in the published protocol.
SmolWAM retains part of the Cosmos Policy model but rebuilds its inference interface with compact visual representations, cached textual features, and twelve transformer blocks instead of twenty-eight. It predicts together actions and the visual evolution of the scene.
After 3.6 GPU-hours of training, the authors announce a 95.5% success rate over 400 LIBERO episodes, with 57 milliseconds per block of sixteen actions and 2.4 GiB of GPU memory. The calculation is eleven times faster than that of the teacher model in their comparison. These benchmark measurements do not guarantee the same performance on all embedded robots.
Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.
Read the full article on the original source
Go to the source →