Robotics · Simulation · Teleoperation
RoboPlayground on Telearms
Physical AI has a data problem. Models need far more real manipulation data than labs can produce, and every hour of it costs hardware, space and someone's time. RoboPlayground turns collecting it into something anyone can do from a browser tab.
Results
- 01Bimanual manipulation missions simulated in MuJoCo, playable in a browser with no hardware
- 02Sub-second control loop, with operator input published on a fixed 20 Hz tick
- 03Four input modalities: keyboard, gamepad, Nintendo Joy-Con over WebHID, Meta Quest over WebXR
- 04Task design grounded in RoboEval: eight bimanual tasks, 3,000+ expert demonstrations
- 05Analytics on peak concurrency, teleoperation hours and per-mission success rate
01
Anyone with a browser
Manipulation research used to need a lab, a GPU and a Python environment. We put it behind a URL, so contributing no longer depends on owning a robot or knowing how to set one up.
02
Play that becomes training data
What an operator experiences as a mini-game is, on the other side of the wire, a reproducible evaluation episode. Every session produces the kind of data these models are actually short of.
03
It has to feel like your own hands
If the arm lags, people stop contributing and the data dries up. Video and control run on separate channels so a congested stream never delays a command, and the loop stays under a second.
04
From a browser to real robots
Simulation is the entry point, not the end of it. The same task set runs on physical bimanual arms, and the same interface drives real hardware, so what people learn here transfers to machines that can break something.