Robotics / world models / in progress
Two robot arms, one shared space
Can two independent arm agents coordinate through a shared, predictive state? What the simulation says, and the plan for the real arms.
- Hardware
- SO-101 arms (LeRobot)
- Simulator
- MuJoCo
- Status
- Simulation done, hardware next
I think the future of world models is agents acting inside a shared reasoning space. This project tests a small, concrete version of that: two robot arms, each run by its own agent in its own process, that coordinate only through a shared, typed and predictive picture of the table. That picture holds each object's status, leased claims on objects and zones, each arm's committed next second of motion, and a handover protocol.
The broad idea isn't new. Systems like VLA-Fabric, LatentToM, CHORUS and Figure's Helix-02 already coordinate arms through shared context or no messages at all. What I haven't found measured is independently written, swappable agents coordinating through an explicit, inspectable shared state over a real network, under latency, dropped links, partner swaps and deliberate corruption of that state.
What the simulation showed
- Across 710 simulated episodes, the shared-state agents matched a central planner on success (100%) with zero arm-to-arm contacts, and finished the sorting task 9.1 ± 2.9 seconds faster.
- They held their speed up to 300 ms of one-way network delay, while the central planner slowed by about 12 seconds between 50 and 300 ms.
- When each arm's goal was private, they succeeded 50 of 50 times against 17 of 50 for the strongest baseline that doesn't communicate.
- Swapping one arm's agent for a differently tuned one worked 18 of 20 times, and exposed a livelock to fix.
Two caveats. These are simulation results, and some cells used only 10 seeds. And today's "world model" is an explicit state plus an analytic forecast of where each arm will sweep, driven by scripted agents. A learned world model comes later in the plan.
Getting to hardware
The hardware bridge is built: per-arm processes over LeRobot, a safety envelope and watchdog, a network bus that can inject delay and partitions, an overhead camera with ArUco markers, and a digital twin that checks for collisions. An audit found 14 safety issues, which were reworked. So far it has only run against simulated servos. The plan is about 20 part-time weeks, ending in a pre-registered study of roughly 300 trials on the real arms.