CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy

2026-06-10 · Source: Artificial Intelligence · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

CHORUS is a novel framework designed for decentralized multi-robot collaboration, leveraging a single Vision-Language-Action (VLA) policy. It addresses the limitations of traditional centralized methods, which scale poorly with team size, and existing decentralized approaches that often require explicit alignment or information sharing during inference. CHORUS adapts a shared VLA backbone to control diverse multi-robot teams, with each robot operating an independent copy conditioned solely on its local observations and a unique robot-identifying prompt. Real-world experiments demonstrate CHORUS's effectiveness, showing a 64% point improvement over decentralized, from-scratch models, a 40% point increase in reactivity to teammate behavior, and superior performance compared to centralized baselines. This indicates that a single VLA backbone can achieve robust decentralized multi-robot collaboration without needing individual robot policies or inter-robot communication at inference time.

Key takeaway

For Robotics Engineers developing multi-robot systems, CHORUS offers a compelling alternative to complex centralized or communication-heavy decentralized approaches. You should consider integrating a single VLA policy to achieve robust, reactive collaboration from local observations. This can simplify deployment and reduce inference-time communication needs. This approach significantly improves task efficiency and reactivity, making it ideal for dynamic environments.

Key insights

A single VLA policy can enable reactive, decentralized multi-robot collaboration using only local observations.

Principles

• VLA models' priors enable reactive collaboration.
• Decentralized control from local observations.
• Shared VLA backbone for diverse teams.

Method

CHORUS adapts a single VLA backbone; each robot runs an independent copy, conditioned on its own observations and a robot-identifying prompt for decentralized control.

In practice

• Improve multi-robot task efficiency.
• Enhance reactivity in team behaviors.
• Reduce communication overhead in robot teams.

Topics

• Multi-robot Collaboration
• Decentralized Control
• Vision-Language-Action Models
• VLA Policy
• Robotics
• Real-world Experiments

Best for: Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• Continuous Reasoning for Vision-Language-Action — Continuous reasoning, shareable and verifiable, bridges the granularity gap between language and continuous robot control.
• Is Diversity All You Need for Scalable Robotic Manipulation? — Not all data diversity benefits robotic manipulation equally; task diversity and debiased expert data are key.
• Network Distributed Multi-Agent Reinforcement Learning for Consensus Control of Quadcopters — ND-MARL enables scalable, communication-aware quadcopter consensus control through a distributed, hierarchical reinforcement learning framework.
• Cooperate to Compete: Strategic Coordination in Multi-Agent Conquest — LM agents can be prompted to improve strategic coordination and negotiation in complex, mixed-motive environments.
• DAM-VLA: Decoupled Asynchronous Multimodal Vision Language Action model — Decoupling temporal processing for vision, language, and action modalities significantly enhances VLA model performance and control robustness.
• LLM-Foraging: Large Language Models for Decentralized Swarm Robot Foraging — Integrating LLMs at key decision points in swarm robotics enhances adaptability and performance without costly re-optimization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.