This site is generated automatically in CI from the runnable policy integrations and benchmark packages that exist today. The home page is comparison-first: use it to cross-check policy status, quality, provenance, and demo coverage quickly, then open a policy folder for charts, Rerun playback, logs, and visual checkpoints.
Each policy now owns its own folder under policies/<policy>/, so the HTML page and raw artifacts live together instead of being scattered at the site root. The same bundle also carries benchmark families under benchmarks/.
Velocity runs use staged locomotion command profiles instead of a single constant command. Reference or pose-tracking cases stay explicitly blocked unless a verified official asset and runtime path actually exist, so the site does not silently drift into mock output.
The public G1 cards currently load a meshless MuJoCo MJCF variant because this repository does not redistribute Unitree's upstream STL mesh bundle. The dynamics stay MuJoCo-backed, while the Rerun robot scene is reconstructed from the same open MJCF kinematic tree.
Serve the generated folder over HTTP for reliable playback. Each policy page lazy-loads the saved .rrd recording and the visual checkpoints inline, so you do not need a second proof-pack HTML flow for normal review.
Successful entries use real checkpoints or public asset bundles cached by CI and must activate the requested MuJoCo transport. The links in each row jump straight to the per-policy folder and raw artifacts. Blocked entries surface the exact missing files or unavailable upstream artifacts instead of falling back to mock output.
| Policy | Status | Quality | Run path | Demo family | Coverage | Ticks | Avg inference | Achieved rate | Dropped frames |
|---|---|---|---|---|---|---|---|---|---|
| GEAR-SONIC gear_sonic Detail page | OK | BAD | REAL MUJOCO SIM | Velocity tracking | Planner-only velocity tracking on the published G1 planner contract | 400 | 4.060 ms | 44.69 Hz | 1 |
| Decoupled WBC decoupled_wbc Detail page | OK | ?? | REAL MUJOCO SIM | Velocity tracking | Lower-body RL locomotion with default upper-body posture | 400 | 0.316 ms | 45.69 Hz | 0 |
| WBC-AGILE wbc_agile Detail page | OK | ?? | REAL MUJOCO SIM | Velocity tracking | Published G1 locomotion checkpoint on the public 29-DOF embodiment | 400 | 0.474 ms | 45.31 Hz | 0 |
| BFM-Zero bfm_zero Detail page | OK | BAD | REAL MUJOCO SIM | Reference / pose tracking | Reference/context walking tracking | 200 | 5.301 ms | 44.96 Hz | 0 |
| GEAR-SONIC Reference Motion gear_sonic_tracking · family gear_sonic | BLOCKED | n/a | REAL MUJOCO SIM | Reference / pose tracking | Published G1 reference-motion tracking with explicit upper-body motion from the official example clip | - | - | - | - |
| HOVER hover | BLOCKED | n/a | REAL SYNTHETIC | Velocity tracking | Multi-modal masked H1 controller | - | - | - | - |
| WholeBodyVLA wholebody_vla | BLOCKED | n/a | EXPERIMENTAL SYNTHETIC | Reference / pose tracking | Experimental KinematicPose contract placeholder | - | - | - | - |
Policies driven by an explicit locomotion command profile. These cards now use the same staged sequence instead of a single constant velocity.
NVIDIA GR00T · Planner-only velocity tracking on the published G1 planner contract
gear_sonic
Real CPU planner_sonic.onnx run inside the MuJoCo-backed G1 showcase, driven by an explicit staged velocity-tracking script instead of a single constant command.
Detail page
Visual checkpoints
Rerun
JSON
log
Proof-pack manifest
Real public GR00T WholeBodyControl run inside the MuJoCo-backed G1 showcase, driven by the same staged locomotion script used for the velocity-only cards.
Detail page
Visual checkpoints
Rerun
JSON
log
Proof-pack manifest
NVIDIA Isaac · Published G1 locomotion checkpoint on the public 29-DOF embodiment
wbc_agile
Real public G1 checkpoint using the published recurrent history tensors and lower-body target mapping, driven by the staged velocity-tracking script rather than a single constant command.
Detail page
Visual checkpoints
Rerun
JSON
log
Proof-pack manifest
Real H1 multi-modal masked policy wrapper for locomotion and body-pose commands, enabled when a user-exported checkpoint is available.
Policies driven by pose targets, motion references, or latent tracking context. If no verified official clip is wired, the card stays blocked instead of inventing a demo.
Real public G1 tracking contract running inside the MuJoCo-backed showcase with a 721D prompt-conditioned observation, IMU gyro/history features, and the shipped walking latent context.
Detail page
Visual checkpoints
Rerun
JSON
log
Proof-pack manifest
NVIDIA GR00T · Published G1 reference-motion tracking with explicit upper-body motion from the official example clip
gear_sonic_tracking · family gear_sonic
Real published GEAR-Sonic encoder+decoder tracking on the official `macarena_001__A545` clip, running inside the MuJoCo-backed G1 showcase with the upstream heading-corrected reference orientation contract.
Experimental AGIBOT X2 kinematic-pose contract wrapper for WholeBodyVLA. The public upstream project does not yet expose a runnable inference release, so this card documents the expected handoff shape and stays blocked until a compatible local model exists.
Normalized benchmark packages that sit beside the policy pages in the same site bundle.
Benchmarks
RoboWBC-vs-official benchmark matrix built from the normalized NVIDIA artifacts.