035 / 200k · A3-fast RKNN · Real-robot teleoperation — Watch the video (3:32).
Explore the 024 / 035 training curves — searchable, interactive TensorBoard scalar snapshots.
Note — Real-robot safety: We strongly recommend using a safety gantry / fall-arrest harness and appropriate protective measures during real-robot teleoperation. These checkpoints do not guarantee stable tracking of every motion; unexpected behavior, loss of balance, and falls can occur. Before playing any motion or deploying your own teleoperation integration on the real robot, first validate it in MuJoCo. MuJoCo simulation cannot fully represent real-robot behavior, and successful simulation does not guarantee safe hardware execution. Keep physical safety protection in place, maintain a safe distance from the robot, and prioritize the personal safety of operators and bystanders.
This project builds on GR00T-WholeBodyControl / SONIC. We sincerely thank the GEAR team for their amazing and solid work, their comprehensive open-source release, and their thorough documentation, which made this A3 adaptation possible.
This repository provides the SONIC whole-body-control workflow for AgiBot A3/A3U, including the 035 and 089 training contracts, MuJoCo simulation, deployment, and real-time XR retargeting.
Start with the end-to-end guide. It provides the authoritative path from A3 flat CSV data to training, sim2sim, ONNX/RKNN, a staged deploy package, and the complete XR teleoperation workflow in simulation and on the real robot.
The identifiers 035 and 089 are experiment labels only; they have no special
meaning and do not denote product versions, model generations, or hardware variants.
- [2026-10-30] Released the 089 / 136k checkpoint with improved sim-to-real performance, standalone AimRT MuJoCo simulation, and a real-time XR retargeting teleoperation pipeline.
- [2026-09-25] Released the initial code and the 035 / 200k checkpoint.
| Model | Training step | Major improvement | Encoders | A3-fast reference | Artifacts |
|---|---|---|---|---|---|
| 089 | 136,000 | 1. More stable upper body, more accurate wrist tracking and a larger workspace for manipulation. 2. Smoother walking with less stumbling and bouncing. |
G1, A3-fast | 10 frames at 20 ms intervals | Hugging Face |
| 035 | 200,000 | G1, A3-fast | 10 frames at 20 ms intervals | Hugging Face |
Both policies control 29 A3 joints at 50 Hz. PT files support training and sim2sim; matching ONNX and RKNN files are used for deployment. Neither checkpoint includes an SMPL encoder.
Download a specific release from the repository root:
python download_from_hf.py --model 089 --component pt onnx rknn
python download_from_hf.py --model 035 --component pt onnx rknnThe release contains the complete A3 adaptation of the SONIC training workflow: Hydra configurations, portable launchers, A3 URDF/MJCF assets, selected flat CSV examples, and the A3-specific reward and observation contracts. The 035 / 200k contract uses a passive-foot model; its PT is hosted on Hugging Face. The 089 / 136k contract uses the later rigid-foot collision model, self-contact rewards and dynamic ankle model. See the 089 checkpoint directory for its artifact and availability details.
Although the 035 checkpoint was trained using the A3 URDF, its domain randomization was specifically tuned to support both A3 and A3 Ultra (A3U). The corresponding models can therefore also run on A3 Ultra.
It includes MuJoCo sim2sim support with the closed-loop ankle/waist serial and parallel joint solver, so the same policy interface can be exercised offline against the A3 model before deployment.
The C++ deployment framework is designed for high-performance inference on A3. It supports cross-compilation for the A3 RK3588S Rockchip board, including NPU inference through RKNN, and a Drive Thor target for A3 Ultra through the Thor runtime variant. The Rockchip / RKNN deployment workflow also supports A3 Ultra (A3U); A3 Ultra can use either the Rockchip deployment path or the Thor runtime variant, depending on the target hardware.
Use the matching G1 and A3-fast ONNX/RKNN artifacts for the selected checkpoint. The 035 step-200,000 artifacts are available from Hugging Face; for 089 step-136,000, follow the artifact preparation and package build steps in the end-to-end guide.
The release includes tested sim-to-real training configurations for A3, including domain randomization of actuator gains, joint friction and armature, body mass, contact materials, joint offsets and external pushes. It also models motor T–N curves and serial/parallel joint torque limits. The 089 configuration adds actuator delay and a dynamic ankle model with randomized motor inertia. See the checkpoint documentation for the 035 and 089 training configurations.
- The 035 and 089 checkpoints provide the G1 encoder and the A3-fast encoder. Neither provides an SMPL encoder or a teleoperation encoder.
- A3-fast is the low-latency deployment path: it uses the G1 encoder with a
20 ms frame interval and exposes future frames through 180 ms. It is not a
separate SMPL or teleoperation representation, so an upstream real-time
retargeting module is still required. This release cannot directly run the
original SONIC SMPL teleoperation chain or the
MotionBricks + VR3 pointhybrid chain. - The current policy is intended for motions whose ground contact stays on both feet or in which one foot is lifted from the ground. It does not currently support kneeling, lying down, standing back up, or other motions that add contact between the ground and additional limbs.
- Fast running and other highly dynamic motions are outside the current capability and validation scope.
- Publish a checkpoint with SMPL and teleoperation encoders.
- Improve sim-to-real transfer and expand abilities, including ground-based motions.
- Add a detailed deployment-on-Thor tutorial for the A3 Ultra Thor runtime.
Source code and the 035 step-200,000 and 089 step-136,000 model artifacts (PT, ONNX and RKNN) are licensed under the Apache License 2.0. Third-party components and data retain their respective licenses and notices; see Third-party and asset notices. The A3 robot description is a modified derivative of the public AgibotTech/A3-A3U-robot-model and is subject to its Mulan Permissive Software License v2 and retained notices.
Real robots can cause injury or damage. Treat the deploy package as a
development reference: first run receive-only and simulator checks, verify
joint mapping and safety limits for the target robot, and use an operator-
controlled staged bring-up. The authoritative MDU service takeover, no-command
probe, and taskset -c 4-5 launch sequence is §6.1 of the
end-to-end guide.