X2 SONIC
Research checkpoints for SONIC whole-body motion tracking on the 29-DOF AgiBot X2 Ultra. This release contains the best checkpoint on the frozen X2 v4 512-motion ruler and the latest fully validated 12-sphere-foot continuation, including its MuJoCo ONNX export.
Read caveat.md before comparing the two checkpoints or using
the MuJoCo result as deployment evidence. The checkpoints have different
training lineages and were evaluated under different termination merges.
Robot URDF/MJCF selection and invariants are summarized in
ASSETS.md. All public model, source, asset, pipeline, and report
URLs are collected in LINKS.md.
Source:
- https://github.com/weikaih04/x2_sonic_training/tree/edc4058
- https://github.com/weikaih04/x2_sonic_data_pipeline/tree/cdbaeb1
- https://dashboard.ai-research-wk.com/reports/humanoid/x2-sonic-current-state-20260731.html
Files
| file | role |
|---|---|
checkpoints/x2-sonic-mainline-50k.pt |
primary checkpoint; best frozen primary eval |
checkpoints/x2-sonic-sphere-44k.pt |
fully gated 12-sphere-foot continuation |
onnx/x2-sonic-sphere-44k-x2.onnx |
X2 policy encoder/decoder export used by the MuJoCo runner |
configs/ |
resolved training and ONNX model configs |
release.json |
hashes, metrics, protocols, and source commits |
caveat.md |
lineage, evaluation, planner, geometry, and sim2real boundaries |
ASSETS.md |
authoritative URDF/MJCF, mass, DOF, head, toe, and collision matrix |
LINKS.md |
public weights, source, assets, pipeline, deployment, and report URLs |
Later sphere checkpoints exist but are not included because they have not completed the same Isaac evaluation and MuJoCo regression.
The primary 50k checkpoint used a G1-permuted warm start and the pre-v4
x2_native_filtered corpus with the X2 reward reform. The sphere 44k checkpoint
is a separate continuation from a mature mesh-foot 42k lineage; its v4 segment
uses official reward values and 12-sphere feet. It is not a later point on the
primary 50k training curve.
Evaluation
The primary checkpoint was evaluated on a deterministic, frozen 512-motion v4 ruler with paper-aligned robot termination filtering:
- success: 480/512 (93.75%)
- mean completion progress: 95.32%
- all-motion MPJPE-L: 47.56 mm
- all-motion MPJPE-G: 249.64 mm
- success-only MPJPE-L: 44.29 mm
MPJPE-L removes global root translation and measures local pose fidelity. MPJPE-G retains global translation and includes trajectory drift. All-motion values are primary because success-only metrics exclude terminated failures.
Sphere 44k measured 450/512 success (87.89%), 91.03% progress, 43.25 mm
all-motion MPJPE-L, and 305.16 mm all-motion MPJPE-G under the released-code
termination merge. That protocol retains foot_pos_xyz and is not directly
comparable to the primary paper-aligned success rate.
The frozen 512-motion ruler is an internal deterministic regression set, not a
SONIC paper test split. Exact active termination terms and thresholds are listed
in caveat.md.
MuJoCo
The sphere 44k ONNX passed a corrected 19-second regression after warmup with zero falls, 3.652 m forward displacement, -0.474 m lateral reference residual, and -1.30 degrees yaw reference residual. Run it from the source repository:
export X2_SONIC_POLICY=/path/to/x2-sonic-sphere-44k-x2.onnx
./run_x2_sonic_mujoco_sphere.sh
The interactive planner is NVIDIA's learned G1 kinematic planner. Online GMR/IK retargets its G1 references to native X2 references before policy inference. The planner itself is not an X2-native planner.
Limitations
- This is Isaac Lab to MuJoCo sim2sim validation, not completed hardware sim2real. Hardware transport, actuator calibration, latency handling, safety supervision, and robot trials are not included.
- The 12-sphere collision model is a controlled approximation rather than a vendor collision asset, and directional drift is not fully removed.
- X2's shorter legs compress G1 planner motion; high jumps often do not produce a clear flight phase.
- MuJoCo armature parity with Isaac Lab is not fully closed.
Terms
The code derives from NVIDIA's Apache-2.0 SONIC release. Robot assets, planner artifacts, motion data, and trained weights may have separate terms. This model bundle is provided for research evaluation; verify all upstream rights before redistribution or commercial use.