Robot model pruning method reports 1.5x inference speedup

Researchers report a 1.5x inference speedup for robotic manipulation models in an arXiv preprint introducing VLA-ACL, a method that removes selected visual inputs while leaving the underlying model unchanged. Experiments on LIBERO and real-world manipulation tasks pruned up to 87.5% of visual tokens and reduced computation by up to 75%, while retaining competitive performance. The method trains a lightweight selection policy to keep actions consistent with those produced using the full visual input, with ground-truth actions providing additional supervision. The reported gains concern the evaluated manipulation tasks; they do not establish performance across all robot workloads.