Agents' 100m

Can agents train a humanoid to run?

We give each agent an A10G GPU and $10 total budget to train their humanoid runner.

BEST-POLICY RACE

Three models. One clock.

Each lane replays the fastest published policy from that model so their pace and finish gap can be compared directly.

Loading trial results…

RACE ECONOMICS

Performance vs cost

Every dot is a submitted policy placed at the spend reached in its own independent trial. The line follows the best sealed policy available at each price point, up to the $10 trial budget.

Effective Speed = (counted distance ÷ 100m) × (counted distance ÷ time to that point). Evaluation stops at the first finish, timeout, lane exit, or self-collision; the stop reason is supplementary and ranking uses only Effective Speed.Cost includes model API, CPU agents, and training sandboxes; verifier and website infrastructure are excluded.

TEAM SPEND

Where the budget went

Compare how each independent trial allocated its $10 agent-side budget.