tinkerbuggy/sonic-x2

SONIC X2 — Whole-Body Control models (AgiBot X2 embodiment)

Deployment models for the X2 humanoid stacks in GR00T-WholeBodyControl-X2, a lean fork of NVlabs/GR00T-WholeBodyControl adding the AgiBot X2 embodiment (31 DOF) with teleop, gamepad, and pkl-playback runtimes.

Contents

file	size	what it is
sonic_policy/x2_sonic_policy.onnx	58 MB	Fused SONIC tracking policy (encoder + decoder, single graph). Input: 1670-D actor observation; output: 31-D joint action (IsaacLab DOF order). Runs in the C++ deploy node (onnxruntime / TensorRT).
sonic_policy/x2_sonic_policy.pt	401 MB	Source PyTorch checkpoint for the ONNX above — for parity checks (eval_x2_mujoco_onnx.py --compare-pt) and re-export.
kplanner_onnx/x2_kplanner_template.onnx	736 MB	Kinematic planner, template mode — torch-free fused export of the motionbricks VQVAE + pose + root pipeline. Consumed by pc2_kplanner_onnx.py. md5-verified byte-identical to the robot-deployed copy.
kplanner_onnx/x2_kplanner_velocity.onnx	734 MB	Kinematic planner, velocity mode (legacy command path). md5-verified byte-identical to the robot-deployed copy.
kplanner_torch/{vqvae,pose,root}/x2_kplanner_*.ckpt	2.2 GB	Source PyTorch checkpoints for the planner pipeline (motion tokenizer / pose model / root model), each with its hparams.yaml, skeleton/, stats/ sidecars — the fine-tuning starting point for the planner.
kplanner_torch/x2_clip.ckpt (+ .modes.json)	1.4 MB	Pose-template library used by template mode.

Layout expected by the code

~/x2_cloud_checkpoints/
  sonic/exported/x2_sonic_policy.onnx     # or pass --model explicitly
  sonic/x2_sonic_policy.pt
  planner_onnx/x2_kplanner_template.onnx  # KPLANNER_ONNX dir

Fetch with:

hf download <repo-id> --local-dir ~/x2_cloud_checkpoints/hf

then point the stack scripts at it (see the GitHub repo's README for the sim_onnx_planner.sh quickstart).

Training provenance (summary)

• Policy: SONIC (universal-token motion-tracking PPO, IsaacLab) trained on large-scale retargeted human/humanoid motion, fine-tuned for soft landing and arm dynamics on the X2 embodiment. MuJoCo sim-to-sim parity verified (ONNX vs .pt max action delta < 1e-5).
• Planner: motionbricks three-stage kinematic planner (motion tokenizer + autoregressive pose model + root model) trained on retargeted locomotion corpora; exported to a single template-conditioned ONNX graph.

License

Weights license under review — currently private / all rights reserved. Robot description and SDK components remain property of their respective owners (AgiBot; NVIDIA for the upstream framework).

SONIC policy (X2) — inference + continue-training kit

• sonic_onnx/x2_sonic_14000_g1.onnx — fused deploy/browser policy (obs 1670 = tokenizer 680 + proprio 990 → action 31, IsaacLab order). md5 ec745672cfefd9507e9d5e9834bf4537.
• sonic_torch/model_step_014000.pt + sonic_torch/config.yaml — the training checkpoint (step 14000) with its as-run config (keep them side-by-side: the config is required for ONNX export and resume).
• training_configs/:
  • sonic_x2_bigrun_nopin.yaml (+ its base sonic_x2_bigrun_smpl.yaml) — the main-run recipe used from ~4k through ~12k+ steps, no fine-tune injection (foot-fixed / nopin corpus).
  • sonic_x2_v7_dance_boxing_drill.yaml — the fine-tune drill: finetune_sample_rate: 0.5 over the v7 slowwalk+dance+boxing set.
  • finetune_rate_schedule.yaml — callback to anneal the fine-tune sample rate without restarting (see in-file notes for the cosine bookkeeping vs a flat 0.5).

Resume training with +resume=true so global_step continues (num_learning_iterations is relative); prefer a fresh run dir to avoid stale meta.yaml wandb re-attachment.