MODEL / CHECKPOINT
ScaleBFM M
Shanghai AI Lab & collaborators
Model
The original ScaleBFM M checkpoint in whole-body mode 7, using the author PyTorch network.
- Organization
- Shanghai AI Lab & collaborators
- Type
- Dense
Continuous-action Humanoid Transformer with cross-attention to future body targets; three history frames and six reference frames. - Version
- M · model22200
author raw PyTorch checkpoint - Robot support
- Unitree G1 · 29 DoF
- Author source revision
- abd6f17c02fe0baabc14709feb8d9ea4959aa621
Checkpoint artifacts
- Author release
- Official ScaleBFM Hugging Face release ↗
- Artifact type
- Author-released checkpoint
- Original file
- checkpoint/humanoid_transformer_m/model_22200.pt ↗
Original raw PyTorch M checkpoint, strict actor/task-embedder loading. No conversion or modified weights; unrelated shared ONNX is not used.
Initialize root history from the first reference root position instead of measured robot position, which is unavailable under the benchmark observation rules. Then use the author mode 7 reference-forcing inference with the original checkpoint; no robot translation is measured or integrated.
Download links deliver the adapter files. The release links and verification notes distinguish author files from conversions; SHA256 identifies the bytes used.
Download ↗
88d5a79946c03ed25503f48b2af71d16290844ef066ca9b6c8fa8dc3837422e3Download ↗
0f09be3a6c5a7396ef7f28ea8a5365b9c723bb896eb4da71cc8809c840dccdecDownload ↗
cc95d2b63a211219b4e5f6911b6bf790c7a33754b81fc293c7f51c3b5e007154Download ↗
3773fff8e969da20c9c76ad82666a08877be04f94676911d913f33e981149f4cDownload ↗
3024f18136314fade2676e22d6eafb1a8b2befa388da55b0885a53fcb1368d16Download ↗
fca11f2399bced32517b51c7f9dcf13c97117ca469b0b32e207d8f7d664fbde9Inference 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
Initialize root history from the first reference root position instead of measured robot position, which is unavailable under the benchmark observation rules. Then use the author mode 7 reference-forcing inference with the original checkpoint; no robot translation is measured or integrated.
Exact checkpoint & evidence ↗- Control interval
- 20 ms
- Reference history
- 0 ms
- Reference preview
- 100 ms
- Output interface
- joint-pd-feedforward
- Configuration
- policy.json