Vision-language-action (VLA) models have shown strong promise in robotic manipulation, yet their post-training still predominantly relies on trajectory-level…
机构:中科院
来源:arXiv 2610.12914 | AI4Papers 论文推荐平台