MuJoCo production capability map
Deployed robotics stack, explicit boundaries, and policy insertion points.
Production-ready for an audited deterministic MuJoCo arm manipulation demo; not yet a trained RL/VLA robotics foundation model stack. Vercel hosts the control/evidence plane; MuJoCo executes inside the Simulation Gateway adapters.
Embodied MuJoCo demos
Language-grounded lift, carry, place, and drop commands with MuJoCo state/action traces, replay, metrics, and artifacts.
Open demoMuJoCo Menagerie and other curated robot models remain provider lanes, not part of the public v0 arm demo.
Open demoLeRobot, MuJoCo MJX/Playground, ManiSkill, and Genesis are explicit provider lanes; dependency checks fail closed until workers are installed and tested.
Open demoMuJoCo engine capabilities
Gateway executes vetted MJCF scenes for deployed adapters. Arbitrary user-uploaded MJCF is intentionally not enabled in production.
Arm cube free-body motion runs through MuJoCo physics with gravity, contact, friction, mass, and release metrics.
The arm lane exposes slide, wrist yaw, and gripper channels through audited actions.
MuJoCo is authoritative for physics. Replay artifacts record native MuJoCo rendering when available; the browser uses state-trace/WebGL replay or deterministic video fallback for inspection.
Every run persists episode JSON, metrics JSON, states JSONL, actions JSONL, result JSON, replay MP4, and manifest artifacts.
The artifact and task graph contracts can carry observations, but explicit MuJoCo sensor blocks are not yet exposed as first-class UI telemetry.
Robotics OS layers
Valid model JSON is treated as the task grounding candidate, then schema validation, bounds, and safety compilation produce executable robot actions.
FastAPI execution plane runs MuJoCo adapters behind the provider contract and returns typed job/artifact metadata.
Deterministic baselines are active; learned policies can replace controllers only behind the same task, action, metric, and artifact contract.
Unsupported Simulation Gateway adapters fail explicitly; the deployed gateway does not advertise or silently substitute mock simulation providers.
Runs already produce synchronized video, state, action, task, and metrics artifacts required for LeRobot-style dataset export.
Long-running RL and VLA training belong on GPU-backed workers or research services, not Vercel request handlers.
Provider registry
mock-compute
submit, status, cancel
mock-simulation
lift-cube
mock-policy
scripted-actions
simulation-gateway / leisaac
LeIsaac is structurally represented but not installed in this scaffold.