AgiBot open-sources SONIC whole-body-control adaptation for A3 humanoid
AgiBot released the AgibotTech/sonic_for_a3 repository, a C++ adaptation of NVIDIA GEAR's SONIC whole-body-control workflow for the A3 humanoid. It includes training code, MuJoCo sim2sim, and RKNN/Thor deployment, with a step-200,000 checkpoint on Hugging Face under Apache 2.0.
China context
- Original name
- 智元机器人
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can fork the repository and adapt the training and deployment pipeline for their own humanoid platforms, using the provided Hydra configs and sim2sim setup as a starting point.
- For investors
- The open-sourcing of a full humanoid control stack under Apache 2.0 may pressure proprietary robotics software vendors and accelerate the availability of low-cost humanoid capabilities.
AgiBot published a focused open-source release adapting the SONIC whole-body-control workflow to its A3 humanoid robot. The repository provides the complete A3 training code adaptation (Hydra configs, launchers, URDF/MJCF assets, CSV examples, reward/observation contracts), MuJoCo sim2sim support, and a high-performance C++ deployment framework targeting the A3's RK3588S Rockchip board via RKNN and A3 Ultra's Drive Thor runtime. It includes a pretrained step-200,000 checkpoint (035) trained on an internal motion dataset for basic walking and simple manipulation, hosted on Hugging Face. The release documents sim-to-real constraints such as T–N motor curves and passive-foot simulation. Current limitations include lack of SMPL/teleoperation encoders, poor wrist tracking, and no support for ground-contact motions like kneeling. Source code and model artifacts are Apache 2.0 licensed; the A3 robot description derives from AgibotTech/A3-A3U-robot-model under Mulan PSL v2. The repository strongly recommends safety gantries and simulation validation before real-robot teleoperation.
The release exposes the full training-to-deployment pipeline for a full-size humanoid: Hydra-based training configs, MuJoCo sim2sim with closed-loop ankle/waist joint solvers, and C++ inference on Rockchip NPU (RKNN) or NVIDIA Thor. The 035 checkpoint uses a G1 encoder with a 20 ms frame interval and 180 ms future-frame exposure, but lacks SMPL/teleoperation encoders, so an upstream retargeting module is required. Passive-foot simulation approximates sole compliance to suppress tiptoe behavior but can introduce instability. The checkpoint's domain randomization supports both A3 and A3 Ultra, enabling cross-hardware deployment.
Developers outside China can now replicate and adapt a full-size humanoid whole-body control stack without access to AgiBot hardware, lowering the barrier to entry for humanoid robotics research and deployment. Competitors in the humanoid space must respond to an open-source, Apache-2.0 licensed pipeline that includes sim-to-real tuning and edge NPU deployment, which can accelerate commoditization of basic locomotion and manipulation skills.
The Apache 2.0 license permits commercial use, enabling startups and enterprises to build on AgiBot's adaptation without licensing fees. The inclusion of RKNN deployment for Rockchip NPUs targets low-cost edge inference, while Thor support addresses higher-performance A3 Ultra deployments. This can reduce development costs for humanoid robot applications in logistics, manufacturing, or service roles.
Observable next signals include: publication of a better wrist-tracking checkpoint, release of SMPL and teleoperation encoders, an open-source teleoperation pipeline, and a Thor deployment tutorial for A3 Ultra. Verification questions: whether the checkpoint's real-robot performance matches simulation, and whether third-party developers successfully deploy on A3 or A3 Ultra hardware.