SuperMimic_Dyn — E009 high-dynamic motion tracking

The final E009, 40,000-iteration checkpoint for a Unitree G1 29-DoF general motion-tracking policy trained on filtered BONES-SEED and LAFAN1 reference motions. Training starts from random initialization and uses a differentiable world model with BPTT actor–critic in MuJoCo/mjlab.

Source and instructions: Baojch/SuperMimic_Dyn, 中文复现与下载指南. The implementation builds on CuboiLeo/SuperMimic-Reproduced, commit d9be385b3eeb13a638fb33fb200d3bfd9be9839f. Thank you to its maintainer for the open implementation and reproduction notes. This is an independent adaptation, not an official SuperMimic release.

Download and contents

Download ckpt_0040000.pt. The source guide includes a download command with SHA-256 verification, data preparation, evaluation and rendering instructions.

• Released file: 221,362,447 bytes.
• SHA-256: 182b7a5d05b6c3e4364d6223d1cf52ba18e6573563e961f6e019d6d445340c3c.
• This is a full PyTorch checkpoint: policy, world model, critic, optimizers, sampler statistics, configuration and progress counters. It is not an ONNX or actor-only package.
• Raw motion data, replay buffers, videos and robot assets are not bundled. Obtain the source repository and appropriately licensed reference motions separately.
• Download only from a trusted source and verify the hash before using the project's checkpoint loaders. The .pt checkpoint uses PyTorch serialization and contains pathlib.PosixPath metadata.

The published checkpoint removes three private machine paths from config.log_root, config.data.library and config.data.sampling_manifest. All 166 tensors and all other configuration values are unchanged. Their equality was checked after serialization; original and released actor outputs were bitwise equal on a fixed CPU batch of four synthetic inputs. No physics evaluation was rerun for this packaging change.

The original local checkpoint had SHA-256 d6fed9645ab4f934dfb3a64d49e90d37509b73b70081b6cc17eeef4bd6ebfe24 and size 221,366,063 bytes. Its historical identity and the new release identity are deliberately distinct; see release.json.

Recorded capabilities and limitations

Evaluation suite	Completed / total	Completion	Mean root error (m)	Mean joint RMSE (rad)
Full validation	3,939 / 3,966	99.3192%	0.07999545	0.20414410
Fixed validation monitor	496 / 512	96.8750%	0.11873723	0.27707667
Training diagnostic	305 / 313	97.4441%	0.15472667	0.35128090
Hard training diagnostic	159 / 164	96.9512%	0.18821980	0.42935862

The suites overlap and their sample counts must not be added. All four reported zero nonfinite clips. Metrics are from the original final checkpoint and use the same released model tensors. See results.json and the source results documentation for definitions and known cross-suite disagreements.

The policy completes training examples of front flips, back flips, side flips, hero landings, cartwheels, crawling and a 168.2-second fall-and-get-up sequence. All 56 canonical flips are training data: 55/56 complete in the training diagnostic and 56/56 in the overlapping hard suite. Validation contains no flip or cartwheel clips, so its overall completion rate does not demonstrate unseen-flip generalization. A validation fall-and-get-up fails after about 13 seconds; a completed crawl has about 0.60 m mean root error. Completion does not establish precise tracking or clean airborne rotation and landing.

This is one seed, evaluated in simulation. The 4,566-clip test split was not evaluated. No real-robot validation or final E009 independent CPU/ONNX closed-loop result is claimed. The fixed 40k endpoint was not selected by validation or test scores.

Model interface

• Robot: g1_no_plastic_hands, removing four hand visual/collision geometries while retaining inertias and the joint tree.
• Control: 50 Hz, four 5 ms physics substeps per action.
• obs: float32 [batch,930], ten frames of 93-dimensional proprioception, stacked term-major. Each frame has relative joint positions (29), joint velocities (29), pelvis-frame angular velocity (3), projected gravity (3), and the previous raw action (29). No measured base linear velocity is included.
• goal: float32 [batch,689], ten 68-dimensional reference frames plus relative torso position (3) and orientation (6). Reference frames contain joint positions/velocities, gravity, angular/linear velocities and height.
• Output: [batch,29] raw actions. The actor applies input scaling internally; the controller converts output using q_target = default_joint_pos + action_scale * action.

Full689 requires actual robot position feedback aligned with the reference coordinate system. Joint encoders and an IMU alone do not directly provide this position. This checkpoint is not a drop-in replacement for an orientation-only deployment policy. Inference requires reference motion input; the policy is not an autonomous motion planner. Exact layout and reset rules are in the source guide.

Training data and budget

E009 uses FlatClean + Ground V4: 103,855 full clips, split into 95,323 training, 3,966 validation and 4,566 test clips. BONES G1 data is resampled from 120 Hz, LAFAN from 30 Hz, to 50 Hz. Metadata filtering excludes object-dependent categories and text, keeps ground and low-posture motions, and allows the exact foam pad label under the remaining filters to retain 56 training flips. Mirrored clips are augmentation. See data.json and the source data documentation for source identity limitations and exact rules.

Training uses two GPUs, 4,096 environments per rank, 40,000 iterations and 327,680,000 real simulation transitions. The recorded run took 4.092 hours including initialization, excluding independent evaluation and rendering. The training configuration is in config.json; data paths are placeholders for external files. Cross-machine bitwise reproduction and identical scores are not guaranteed.

License and data notices

The implementation's MIT license is not asserted as the license for these trained weights. Public availability does not constitute a grant of unrestricted commercial-use or redistribution rights. This notice does not grant rights beyond those held by the publisher or replace applicable third-party terms.

Training data includes Motion Data by Bones Studio. Use of the underlying dataset is subject to the BONES Motion Capture Dataset License Agreement. Its eligibility, attribution, results-distribution and use restrictions remain applicable; this model card does not certify their satisfaction for any downstream use.

Training also used LAFAN1, provided under CC BY-NC-ND 4.0, through lvhaidong's G1 retargeting. Data preparation included retargeting, filtering and resampling. The source license does not expressly determine the licensing status of trained weights; this notice does not resolve whether these weights constitute Adapted Material. The noncommercial/no-derivatives source terms must not be represented as an unrestricted weight license.

Implementation based on CuboiLeo/SuperMimic-Reproduced, copyright © 2026 Yuxi Liu. The source repository retains its MIT notice and third-party notices. Please cite the SuperMimic method paper, the implementation baseline and the datasets, including Harvey, Yurick, Nowrouzezahrai and Pal, Robust Motion In-Betweening, ACM TOG, 2020.