CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy

ListenOne Brain, Many Robots: The CHORUS Trick9:45

Show notes

CHORUS is a single vision-language-action policy, fine-tuned from π0.5, that drives whole teams of robots — but each robot runs its own independent copy, sees only its own cameras, and never talks to its teammates. We dig into how a shared backbone plus a robot-identity prompt gets coordination "for free," why the decentralized version beats a centralized one, and what the recipe suggests for anyone building large pretrained models. As always, we flag what's a real result versus a project-page claim from an under-review paper. **Sources** - CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy — Ria Doshi, Tian Gao, Annie Chen, Chelsea Finn, Jeannette Bohg

Subscribe to the channel: https://podcast.lumalabs.link/27a25f51ea81d41a8253f46c6133a35e/feed.xml

TL;DR

CHORUS (Stanford; Doshi, Gao, Chen, Finn, Bohg) is a single VLA policy fine-tuned from π0.5 that drives multi-robot teams via fully decentralized deployment: at inference each robot runs an independent copy conditioned only on its own cameras plus a short robot-identifying prompt, with no shared observations, no proprioceptive channels, and no inter-robot communication. The training recipe samples single-robot (observation, action) tuples from multi-robot demonstrations and predicts a padded 32-dimensional action so one weight set can drive embodiments with different action spaces and control rates. On three real-world tasks (laundry-basket lifting, tape-measure extension, book handover) across Kinova, ARX, and YAM manipulators, CHORUS reports a 64-point mean-success gain over per-robot from-scratch diffusion policies, +40 points in reactivity to teammate perturbations, and — most surprisingly — beats a centralized VLA that sees the full team’s observations, at 90% success on a three-robot basket-transport task.

Key claims

A single VLA backbone fine-tuned on multi-robot data and deployed independently per robot achieves decentralized multi-robot collaboration without alignment procedures or inference-time communication [§Method].
A shared 32-dimensional padded action space plus a per-robot identity prompt lets one policy drive multiple embodiments with different action spaces and control rates [§Method].
CHORUS beats decentralized from-scratch diffusion policies by 64 percentage points in mean success rate across the real-world task suite [§Results, laundry-basket / tape / handover].
Per-robot from-scratch diffusion policies exhibit a characteristic “mismatch” failure — one robot proceeds before the other has secured its handle — that vanishes when a single policy is trained on all robots’ perspectives [§Results, laundry-basket qualitative].
Under teammate perturbations, CHORUS recovers 40 points more often than a backbone fine-tuned per-robot, reacting to teammate behavior nearly 2× more effectively [§Reactivity].
CHORUS outperforms a centralized VLA that conditions on the entire team’s observations despite conditioning on strictly less information; the paper attributes this to the centralized inputs breaking the backbone’s pretraining semantic correspondence [§Centralized comparison].
The same recipe scales to a three-robot Kinova+YAM team reaching 90% success on collaborative basket transport with no architectural change and constant per-robot context [§Three-robot].

Method

Each robot in a multi-robot demo contributes single-robot (observation, action) tuples to a shared training buffer. A single VLA backbone (π0.5) is fine-tuned on this buffer with a robot-identifying prompt prepended at every timestep (so the same weights condition on “which robot am I”), and the action head predicts a 32-dimensional padded action vector so heterogeneous embodiments (Kinova mobile bimanual, ARX arm, YAM mobile) with different DoFs and control rates map into one output space. At deployment, an independent copy of the shared weights runs on each robot, taking only that robot’s cameras plus the identity prompt — no cross-robot state, no communication channel, no aligned team observations. Because each robot acts on its own observations, execution is asynchronous and the context window stays constant as team size grows.

Results

Laundry-basket lift (two-robot): CHORUS succeeds where per-robot decentralized diffusion policies collapse to mismatched-timing failures; +64 pp mean success gain vs the from-scratch decentralized baseline across the task suite.
Tape-measure extension (Kinova + mobile YAM at different control rates): successful tightly-coupled cross-embodiment interaction, including a variant with distractors.
Book handover: precise spatial coordination at the moment of exchange, purely from local views.
Teammate-perturbation reactivity: +40 pp recovery rate vs a per-robot fine-tuned backbone, ~2× more reactive to teammate behavior.
Centralized comparison: CHORUS (decentralized) beats the centralized VLA baseline in mean success — an inversion of the usual centralized-upper-bound expectation, attributed to input distribution matching the backbone’s pretraining.
Three-robot scaling: single policy on Kinova+YAM three-robot team reaches 90% task success on collaborative basket transport with no architectural change.

No numerical benchmark table beyond these percentage-point summaries is quoted on the project page; all videos are shown at 4× speed and the paper is listed as “under review” for 2026.

Why it’s interesting

CHORUS is the sharpest counter-recipe filed so far to the “decentralized multi-robot requires per-robot policies + explicit alignment / communication” default, and it lands squarely inside the VLA Models design debate: it argues that the visuomotor prior of a pretrained VLA is strong enough to encode teammate-modeling as a byproduct of training on multi-robot data, without any team-level architecture. Two connections are worth flagging. First, it complements Introducing Waddle: Agents that Control Robots‘s multi-robot claim from the opposite direction: Waddle solves multi-robot with an LLM-agent-as-top-level-policy that spawns coordinating subagents over frozen VLAs; CHORUS keeps the VLA as the top-level policy and denies the orchestrator entirely, arguing coordination emerges from shared weights + per-robot identity conditioning. Second, its finding that a centralized VLA underperforms the decentralized version because the multi-robot observation stack breaks the backbone’s pretraining distribution is a concrete instance of the broader observation on this concept page that matching the backbone’s pretraining input distribution is often the load-bearing lever — sibling in spirit to See like a Robot: Robot-Centric Pointmaps for Vision-Language-Action Models‘s “preserve the H×W token grid the pretrained VLA expects” argument for viewpoint robustness. Load-bearing weakness: no benchmark table is public at filing, so the 64-pp / 40-pp / 90% numbers are project-page claims from an under-review paper rather than reproduced results.