Reproduce “MorphAct: Robot Policies Morphing with the Worlds They Act In” on OpenResearch

About this paper

Ziyan Liu, Jiting Liu, Weihong Zhong, Wenda Yu, Yeqiu Chen, Jason Li, Xiaojiang Peng, Shuyuan Hu and 2 more

MorphAct adapts a robot policy to its current interaction context by generating low-rank updates for vision-language and action modules. The action-update generator uses features from the adapted vision-language module, linking perception and action adaptation. In an X-VLA LIBERO ablation, this coupling reached 97.20% mean success versus 86.00% without it; physical gains were measured on four tasks.

Key insights

• At the analyzed checkpoint, negative gradient alignment appeared in 47.80% of cross-task comparisons for the vision–language module and 48.28% for the action module, indicating local task interference.
• Context-conditioned generation outperformed fixed or task-ID-based conditions in the reported ablations; using adapted rather than pre-adaptation features to generate action updates further improved results.
• The approach trades additional inference cost for performance: reported chunk latency increased by roughly 18–21% relative to shared LoRA, alongside higher memory use.

Summary by alphaXiv.