From private job to Arena
How comparison, evaluator trust, lineage, and the public learning loop work.
How Arena works
- Open a challenge and read the task, trial count, and scoring method.
- Package the policy entrypoint from your existing repository.
- Submit the policy you want to evaluate.
- HIM runs it with the same robot, starting conditions, trials, and evaluator used for every entry.
- See the score, builder, policy version, and evidence behind the run.
- The winning eligible Table Tennis policy may be reviewed for a supervised session on HIM’s real Unitree G1.
A visible path from simulation to a real robot
Every HIM Official challenge shows the task, the reward, and the steps between a simulation score and hardware. Table Tennis leads to a supervised sim-to-real session on HIM’s Unitree G1. The completed Nick Kosir Dance challenge shows a prior full path: Stanford students at HIM retargeted the reference motion, trained and stabilized the policy in simulation, then transferred it to Adam.
- Table Tennis is open: rank under the shared evaluator, pass HIM’s hardware review, and earn the winning eligible policy’s supervised Unitree G1 session.
- Nick Kosir Dance is complete: more than 150 million training iterations led from reference video through MuJoCo and Isaac Lab to Adam.
- Skateboarding is upcoming: test balance, recovery, and whole-body locomotion on a moving board.
Build with the challenge data
HIM's real Rokoko motion, Orbbec RGB-D, and WitMotion paddle IMU table-tennis captures are available to download now. Hidden evaluator episodes stay private so the leaderboard measures generalization instead of memorization.
- Released challenge data can be used with any training code or simulator.
- Earned hardware data gives active builders access to hard-to-collect robot trajectories, telemetry, and synchronized video.
- Hidden evaluation data is isolated from candidate code and never bundled into a public starter set.
What makes a score official
- The evaluator version and digest are fixed before the result is published.
- Candidate output freezes before evaluator execution.
- Only evaluator-attested numeric metrics become authoritative.
- The artifact manifest and result attestation bind the run, package, environment, and provider job.
- Blind battle identity stays hidden until the vote/reveal rule permits it.
- Hardware evidence requires a separate supervised Unitree G1 review; simulation never silently becomes a hardware claim.