Skip to content

Releases: NJU-R-L-Group-Embodied-Lab/SuperMimic-Repro-Sim2Real

v1.1: 闪身步 Pro (e24_0355000, spin-and-landing dance)

Choose a tag to compare

@Darkness-hy Darkness-hy released this 27 Oct 09:18

闪身步 Pro / spin-and-landing dance on the real G1.

e24_0355000 ran shanshen_duosuomao_lean35kfs_pad2 (21.9 s, two fast spins and a jump with a one-foot landing) from the first frame to the last on our real Unitree G1, in two runs in a row (ground stage, safety rope attached). Before it, every real run of 9 earlier checkpoints on this dance aborted. One checkpoint, two runs, no ablation: see docs/REAL_ROBOT_LEAN35KFS.md for the failures, the fixes, which sim gates predict the real robot, and the deploy command.

Assets

  • e24_0355000_lean35kfs.tar.gz: extract at the repository root. Contains deploy_artifacts/e24_0355000/ (policy.onnx is ONNX IR 9 and runs on PC2 as is, policy_ir10.onnx, deploy_config.json, SHA256SUMS), deploy_artifacts/ckpts/e24_0355000.pt, the clip goal table deploy_artifacts/clips/shanshen_duosuomao_lean35kfs_pad2.npz, the training clip / library / real-reset states under data/, and the runner log of the first successful run under deploy_logs/real/. This model is fine-tuned on this one clip only; it is not a general tracker.
  • shanshenbu_pro_real_robot.mp4: demo video (left: human source video, right: real G1).
  • SHA256SUMS: checksums of all release files (v1.0 and v1.1).

Training recipe: scripts/train/train_lean35kfs_e24.sh (its parent checkpoint is not released). Sim2sim gate: scripts/train/gate_lean35kfs.sh. All new training options are opt-in and off by default.

Safety: the real-robot code drives the motors directly. Follow docs/DEPLOY_G1.md, keep the rope attached and the remote in hand. Use at your own risk.

v1.0: pretrained models and real-robot demo

Choose a tag to compare

@Darkness-hy Darkness-hy released this 25 Oct 20:31

Pretrained checkpoints and the real-robot demo video for SuperMimic-Repro-Sim2Real.

File What it is
e11c_0220000.tar.gz Base model trained on BONES-SEED (paper observation). Training checkpoint + ONNX deploy bundle.
e13i_0291000.tar.gz Fine-tuned from the base model on BONES-SEED + the 4 custom dances in motions/custom_dance/. This model ran the ~46 s dance on a real G1.
shanshenbu_real_robot_v3.mp4 Real-robot demo video (闪身步 / shanshenbu), human source video on the left, real G1 on the right, ~46 s, 720p.
SHA256SUMS Checksums.

Each tarball contains ckpt_*.pt, policy.onnx, policy_ir9.onnx (for the older onnxruntime on the robot's onboard PC2) and deploy_config.json.

Usage: tar -xzf e13i_0291000.tar.gz -C deploy_artifacts/, re-run the sim2sim gates (docs/SIM2SIM.md), then follow docs/DEPLOY_G1.md. Always run on a hoist with someone holding the remote (L2+B = damping). Use at your own risk.

Both models were trained on BONES-SEED; check its data terms before use. The code builds on CuboiLeo/SuperMimic-Reproduced (Yuxi Liu, MIT).