Same parts, every team. Twist-on legs, hand-tight only — the difficulty floor is bimanual coordination and 3D spatial reasoning, not fasteners.
Teams compete on the AI policy driving the G1 — no teleop, no human steering. Each run is judged on outcome and time, against a human reference run on the same parts and scene.
Three resources released before the event: a real-world dataset, a high-fidelity simulation environment, and remote-lab access for policies that need real-robot validation.
A shared dataset of teleoperated G1 episodes performing the UTTER table assembly task. Use it to bootstrap policies before sim or remote-lab evaluation.
A high-fidelity simulation environment identical to the venue setup. Train policies, iterate on contact-rich behaviors, and evaluate against the same Inventory · Partial · Final checkpoints used on the floor.
Teams without on-site G1 hardware submit policy checkpoints to organizer-hosted G1s and receive a real-world score report before the venue. Lowers the hardware barrier without changing the rules.
Tell us about your team. We'll be in touch with simulation access, the dataset link, and the call-for-teams when it opens.
Open the form →