Benchmark

The Physical AI Control Benchmark

A public, open benchmark of embodied control — raw speed on real physics, every ranked run verified on the server. But note what a time-trial rewards: the policy that commits fastest, not the one that holds its own boundary cheapest. The Institute's load-bearing metric is J/VT — joules per viability-second held under disturbance — the axis on which a certificate-gated agent beats a reckless policy (see the Agency Ladder). These speed boards are one axis; agency is the other. Standings and the full API are open and citable.

How a run is verified

There is no self-reported score. When a run finishes, the client submits the whole trajectory, and the server replays it: it re-checks that the course was completed in order, recomputes the authoritative time or distance from the trajectory itself, and rejects any segment that moves faster than physics allows (anti-teleport). The value the client claims is ignored.

PhysicsMuJoCo, deterministic per course.

CourseFixed per task (course id c1). Same gates, same order, for everyone.

AuthorityServer recomputes the score from the submitted trajectory.

Anti-cheatPer-step segment-speed cap. Teleported runs are refused.

LicenseStandings are open data under CC-BY-4.0.

The tasks

Task	Metric	Board id
🚗 Rover drive course	time to finish (s, lower is better)	`rover`
🚁 FPV drone ring time-trial	time to finish (s, lower is better)	`fly`
🦾 Robot-arm fastest stack	grip-to-stacked time (s, lower is better)	`arm`
🌊 Underwater AUV marker course	time to finish (s, lower is better)	`auv`
🐕 Quadruped waypoint walk	time to finish (s, lower is better)	`quad`
🚶 Humanoid waypoint walk	time to finish (s, lower is better)	`humanoid`
🧠 Forge learning trials	learned-gait distance (m, higher is better)	`forge`

Live standings

Pulled live from the public API below. The fastest run on each embodiment, worldwide.

Loading standings…

Query it

One public, versioned, CORS-enabled endpoint returns every board. Fetch it from anywhere.

GET https://physicalai-bmi.org/api/leaderboards

GET https://physicalai-bmi.org/api/leaderboards/{board_id}

$ curl https://physicalai-bmi.org/api/leaderboards
{
  "schema_version": "1.0",
  "benchmark": "Physical AI Control Benchmark",
  "license": "CC-BY-4.0",
  "verification": "Server-side trajectory replay ...",
  "boards": [
    { "id": "arm", "task": "Robot-arm fastest stack",
      "value_field": "time", "unit": "s", "order": "asc",
      "count": 12, "runs": [ { "name": "...", "time": 4.03, "ts": 0 } ] }
  ]
}

Schema 1.0. value_field names the run key to rank by (time ascending, or dist descending). CORS open. Cached 60s.

Cite it

The benchmark, its methodology, and its standings are open. If you use them, cite:

@misc{physicalai_control_benchmark,
  title  = {Physical AI Control Benchmark},
  author = {{Institute for Physical AI}},
  year   = {2026},
  howpublished = {\url{https://physicalai-bmi.org/leaderboards}},
  note   = {Server-verified embodied-control leaderboards}
}

Institute for Physical AI. Physical AI Control Benchmark. 2026. https://physicalai-bmi.org/leaderboards