MODEL / CHECKPOINT
BFM-Zero
CMU / Meta
Model
Forward-backward behavioral foundation model, using the author's zero-shot motion-tracking inference.
- Organization
- CMU / Meta
- Type
- FB
Forward-backward latent policy with a backward reference encoder. - Version
- Author release · 62b4206
reproducible export from author checkpoint - Robot support
- Unitree G1 · 29 DoF
- Author source revision
- df9c4a9633aa155c727419e82c906010b1e1845f
Checkpoint artifacts
- Author release
- LeCAR-Lab/BFM-Zero ↗
- Artifact type
- Reproducible export from author checkpoint
conversion/prepare.py verifies the original artifact and source hashes, strictly loads62 inference tensors and exports a hash-locked ONNX. Fresh macOS export matches prior Linux SHA256.
Download links deliver the adapter files. The release links and verification notes distinguish author files from conversions; SHA256 identifies the bytes used.
Reproduce export ↗
cdc9d6454d72b00fabfd1c2ab26ad463a5638cde4e3ac82946975d1e459d9939Inference interface
The adapter declares its control clock and reference window below. Source and configuration describe the observation and action transforms.
Observation requirements
Model-specific handling
reference_privileged is a REFERENCE-only backward-encoder input. Actor proprioception and the author double history insertion are retained; no measured translation.
Exact checkpoint & evidence ↗- Control interval
- 20 ms
- Reference history
- 20 ms
- Reference preview
- 140 ms
- Output interface
- joint-pd-feedforward
- Configuration
- policy.json