Reproduce “VLA-MOPD: On-Policy Distillation for Multi-Task VLA Policy Consolidation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yongxuan Lv, Zhen Fang, Lichen Man, Hao Li, Pu Li, Hengyi Wang, Langfeng Cao, Lin Chen and 5 more
VLA-MOPD consolidates task specialists into one robot policy by letting the student act and using instruction-matched frozen teachers to guide intermediate states in its action-generation process. The student is trained on observations from its own rollouts, rather than only fixed demonstrations. In LIBERO, its average success-once rate reaches 98.9%, versus 98.15% for endpoint action distillation, in a one-seed evaluation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.