MODEL / CHECKPOINT
HEFT PMG
Tsinghua / RobotEra / SQZ
Model
Checkpoint artifacts and an independent MotionChallenge inference adapter.
- Organization
- Tsinghua / RobotEra / SQZ
- Type
- Dense
Dense policy with fully connected layers and normalized hidden features. - Version
- SHA 192aaacc
checkpoint SHA256 - Robot support
- Unitree G1 · 29 DoF
- Author source revision
- 0d5ba31e33397f3543d350d98b637e26d92f470a
Checkpoint artifacts
- Author release
- Axellwppr/motion_tracking ↗
- Artifact type
- Author weights with verified input wrapper
- Original file
- G1_PMG/policy.onnx + policy.onnx.data ↗
The distributed single-file ONNX adds one input Concat. All 33 initializer tensors and all 29 native graph nodes are preserved exactly; action outputs matched on 100 random input cases (maximum absolute error 0). This checks the checkpoint wrapper, not full controller/reset parity.
Checkpoint comparison ↗Download links deliver the adapter files. The release links and verification notes distinguish author files from conversions; SHA256 identifies the bytes used.
Download ↗
192aaacc49ce066b85bdc9822bc35eca9ae21b8efc8371c089f7bc0077a86791Inference 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
Released actor uses joint state, angular velocity/gravity or orientation and reference features. Retain author normalization, history/recurrent state, action processing and declared lookahead; validate numerical independence from actual translation.
Exact checkpoint & evidence ↗- Control interval
- 20 ms
- Reference history
- 320 ms
- Reference preview
- 120 ms
- Output interface
- joint-pd-feedforward
- Configuration
- policy.json