VLAICLR 20252025

VLA-Transfer: Cross-Embodiment Transfer for Vision-Language-Action

Yao Mu, Bingxian Wu, Ping Luo, Chelsea FinnUniversity of Hong Kong, Stanford University

摘要

We present VLA-Transfer, a framework for cross-embodiment transfer of vision-language-action policies. VLA-Transfer enables policies trained on one robot to transfer to different robot morphologies with minimal adaptation, leveraging shared visual-language representations across embodiments.

cross-embodiment transferVLApolicy transfershared representationminimal adaptation

技术细节

数据集
Open X-Embodiment (13 robots)DROID
仿真平台
测试机器人
模型骨架

Cross-Embodiment Transformer

编码器

Multi-View ViT + Embodiment Token

解码器

Universal Action Decoder (6 layers)

京ICP备2026064258号-1