Ep#93: CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy
Bria, a first-year PhD student, and Janette Boke, an associate professor at Stanford, discuss CHORUS, a decentralized multi-robot collaboration policy, with RoboPapers host Michael.
In this talk
- 2:27
Why robots need to collaborate
Bria uses apartment moving, construction, and home cleaning to show how teams can handle tasks that are difficult or slow for one robot. She frames CHORUS as an early step toward teams with different robot bodies and roles.
- 6:35
Designing for local, independent control
The authors explain that centralized policies need camera views and state from every robot, which adds communication needs, latency, and a growing observation space. CHORUS instead uses each robot’s local camera views and proprioception, without sharing proprioceptive state. The team must make sure a robot can see the other robots when their actions matter.
- 20:14
One VLA policy across different robots
CHORUS fine-tunes a pretrained π0.5 VLA with LoRA on data from robots with different control frequencies, sampling examples to balance their representation. Each robot gets its own policy copy at deployment, conditioned on its own views and a role identity prompt. Frequency-aware action chunk sizes help align robots operating at different rates.
- 26:15
Real-world results and scaling
Across real-world basket lifting, tape measuring, book handover, and a three-robot laundry basket task, pretrained CHORUS outperformed a decentralized diffusion baseline by 64 percentage points. Weight sharing improved performance in a 20-trial test where one robot had to track a moving book. In the three-robot task, 30 demonstrations taught two basket-carrying robots to wait and adjust based on what another robot did.
- 38:41
Limits and next steps
The authors say simulation did not work as well as hoped, partly because pretrained policies may not transfer well to simulation data in low-data settings. They identify teleoperation and multi-robot data collection as major bottlenecks. Future directions include better simulation, more efficient data collection, and ways for robots to exchange information while remaining decentralized.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.