BFM-Zero

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. This page keeps the full charts, playback, and downloadable artifacts.

Generated: 2026-04-24 02:27:15Z Commit: f1b940ac0f78 Actions run

BFM-Zero

CMU · Reference/context walking tracking

bfm_zero

BAD REAL MUJOCO SIM MOTION_TOKENS runtime.motion_tokens

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.

Robotunitree_g1

Ticks200

Avg inference5.301 ms

Achieved rate44.96 Hz

Target positions

Joint 0 actual vs target

Observed joint velocity

Inference latency

Embedded Rerun viewer Fetches run.rrd lazily when the viewer enters the viewport.

Preparing interactive view Loads the viewer runtime and recording on demand when visible.

Case key bfm_zero

Policy family bfm_zero

Command kind motion_tokens

Expected behavior Reference/context walking tracking

Showcase transport MuJoCo sim

Embodiment assets/robots/unitree_g1/g1_29dof.xml

MuJoCo model variant meshless-public-mjcf+ground-plane

Command data [0.000]

Checkpoint source Prepared BFM-Zero ONNX checkpoint plus tracking context assets

Demo family Reference / pose tracking

Demo sequence Replays the shipped `zs_walking.npy` latent walking context. No verified public waving or upper-body mocap clip is bundled in this repo today.

Model artifact models/bfm_zero/bfm_zero_g1.onnx

Command source runtime.motion_tokens

First target frame [-0.174, -0.006, -0.025, 0.414, 0.260, 0.013, ...]

Last target frame [-0.145, 0.184, 0.044, 0.049, -0.369, 0.215, ...]

Quality verdict BAD

Verdict basis achieved rate 45.0 Hz < 45

Mean joint error 0.130 rad

Joint error p95 0.422 rad

Peak joint error 2.553 rad

Min base height 0.747 m

Frames height < 0.4 m 0

Rerun recording · JSON summary · run log · Proof-pack manifest · configs/bfm_zero_g1.toml

Visual checkpoints

Each image overlays the target pose in blue against the actual replayed pose in orange. The checkpoints are selected from the replay trace so you can cross-check startup, motion onset, peak latency, furthest progress, and final state without opening a second report.