EmbodiWeave

Turn robot failures into evidence-backed training decisions

Simulated world Recorded Graph … Worker …
Watch the recorded MuJoCo physics comparison

Actual recorded contact-physics states, with the same seed on both sides. Both checkpoints were trained in 2D and run unchanged in MuJoCo. This is a selected illustration; all transfer results, including failures, are reported.

Watch mixed-domain MuJoCo development (recorded)

One selected development episode from the 2D release and graph-ranked physics lineage, on the same seed. The right checkpoint trained in both simulators. Its full results and remaining crossing limit are on the physics evidence page. No physics confirmation was opened.

Arena replay

Positions come from recorded simulator states of the selected evaluations
Before
After
t=0
robotmoving obstacle static obstaclegoal ✓ success✕ collision◷ timeout

Scenario controls

Runs a real evaluation of a saved checkpoint on the configuration you set

Candidates all eligible checkpoints, ranked by rule

Rejected

    What-if (isolated graph copy)

    Interventions go into a GRAPH.COPY of production. Results are hypothetical and cannot launch training.