Event
CoEvolve: A Self-Evolving Code Framework for Humanoid Robot Collaboration
1 reports · 1 sources
Zicheng Liu et al. propose the CoEvolve framework, which represents multi-robot collaboration as executable code built on top of fixed motion skills, with verifier-guided co-evolution revising the policies: physical verification evaluates each execution transition, and rejected behaviors trigger structured revisions from accepted checkpoints. In a task where two simulated Unitree G1 humanoid robots carry a rigid payload through a confined corridor, across 15 base-task runs covering three types of multimodal foundation models, CoEvolve completed 8 tasks, achieving end-to-end success with each type of model, but the paper also notes that persistent limitations remain in motion precision and contact stability. The work is currently under review at ICLR 2027.
Report timeline
- CoEvolve: A Self-Evolving Code Policy Framework for Humanoid Robots
CoEvolve is a framework that represents team coordination as executable code over fixed robot motor skills. Through verifier-guided co-evolution, multimodal foundation models instantiate coordination strategies, physical verification evaluates each executed transition, and rejected behaviors trigger structured revisions starting from an accepted checkpoint.
arXiv cs.RO ·