CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy

Summary

CHORUS fine-tunes one pretrained VLA policy to run independently on each robot in a heterogeneous team. Each robot uses only its own observations and an identity prompt, yet the shared policy coordinates real mobile manipulators without runtime communication.

Problem

Multi-robot manipulation needs robots to respond to teammates while acting under partial observability.
Centralized policies require team-wide observations and actions, so context length and communication needs grow with team size.
Decentralized per-robot policies can require shared cameras, teammate proprioception, online alignment, or one trained policy per robot, which makes deployment and scaling harder.

Approach

CHORUS starts from the pretrained $\pi_{0.5}$ vision-language-action backbone and fine-tunes it with LoRA on multi-robot demonstrations.
The training data is split into single-robot tuples $(o_r^t, A_r^t, c_r)$ , so the model sees one robot's local observation, action chunk, and identity prompt at a time.
A robot-identifying prompt names the embodiment and role, allowing one shared policy to choose actions for different robots.
Padded 32-dimensional action vectors and variable image tokens let the same policy handle different robots, sensors, and action spaces.
At inference, each robot runs its own copy of CHORUS using only local observations; no inter-robot messages, shared cameras, or shared proprioceptive state are used.

Results

In real-world tests on basket lifting, tape measurement, book handover, and 3-robot transport, the paper reports 25-45 demonstrations per task and 10-18 evaluation rollouts per task.
CHORUS improves mean task success by 64 percentage points over decentralized diffusion policies trained from scratch.
In a teammate-perturbation handover test, CHORUS succeeds in 17/20 trials versus 9/20 for the same VLA backbone trained as separate per-robot policies, a 40 percentage point gain.
The perturbation split is 8/10 versus 3/10 for left perturbations and 9/10 versus 6/10 for right perturbations.
CHORUS matches or exceeds a centralized VLA baseline across the three 2-robot tasks, although the centralized baseline conditions on both robots' observations.
On a 3-robot transport task with Kinova and YAM mobile manipulators, one unchanged CHORUS policy reaches 90% success.

Link

https://arxiv.org/abs/2606.12352v1