Separate study / mixed 2D and MuJoCo training

Train for the physics.
Keep the evidence.

Unchanged 2D policies exposed a gap in MuJoCo. This separate, preregistered study trains in both simulators. FalkorDB ranks the origins using four original 2D validation suites and four physics development suites.

Loading the recorded study result…

Recorded MuJoCo development / same seed

See what changed in physics.

Left: the confirmed 2D release, transferred unchanged. Right: the graph-ranked lineage trained in both simulators. This selected development episode shows a pedestrian collision and a goal. The full suites below include every outcome; crossing still missed the extra clearance rule.

Seed 6403036. Recorded poses are sampled and held without interpolation. Titles, camera movement, rays and trails are render overlays. This is not physics confirmation or a hardware test.

Validation / completed trials

Every attempt stays visible.

The original three trials each use seed 23 and request 300,000 additional PPO steps. One separately preregistered longer follow-up retains the selected parent’s recipe and seed, with a 600,000-step budget. The original promo-1 gate is unchanged. Selection requires all eight validation suites and an extra episode of collision clearance on each physics suite. The table shows physics development collisions, with its per-suite limit. Pass also requires the original goal-success floor: target 80%, open 90%, static 75% and crossing 75%.

Loading measured development records…

One protected batch / physics only

Confirmation follows selection.

One hundred fresh MuJoCo episodes per family, opened only after the candidate hash is frozen. These outcomes never enter the planning graph. No new 2D held-out evaluation was authorized for the physics child; the confirmed 2D parent remains the default 2D demo policy.

Loading confirmation status…

Live FalkorDB / development evidence only

Why this origin?

The authored rule ranks failed suites, collision clearance and measured deficits. Each Decision saves the exact Cypher, parameters and returned records before training. The returned promotion verdict refers to the original 2D gate; physics selection is reported separately.

Ready to query FalkorDB.

Exact query receipt

Download the measured index · Protocols, trajectories and checksums