Reproduce “Co-VLA: Coordination-Aware Structured Action Modeling for Dual-Arm Vision-Language-Action Systems” 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
Yandong Wang, Jiaqian Yu, Xiongfeng Peng, Lu Xu, Yamin Mao, Weiming Li, Jaewook Yoo, Dongwook Lee and 3 more
Researchers from Donghua University and Samsung R&D Institute China-Beijing developed Co-VLA, a framework that introduces explicit structural priors into Vision-Language-Action (VLA) models for dual-arm robot systems. This framework enhances the reliability and interpretability of bimanual manipulation by decomposing actions into shared coordination and arm-specific residuals, achieving up to 91% success in a simulated "Handover Block" task and 73% in real-world handovers.
Key insights
- Bimanual collaboration is best modeled as a structure over actions, rather than a single monolithic action, providing a more interpretable and robust inductive bias for VLA systems.
- Explicitly separating shared coordination intent from arm-specific execution details, coupled with task-adaptive losses, improves learning of varied bimanual coordination regimes.
- An adaptive, latent-aware controller operating at execution time can refine robot movements, addressing high-frequency noise while preserving critical micro-adjustments, without requiring specialized hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.