A humanoid robot's C++ nervous system, open-sourced

What it does

agibot_x1_infer is the control and inference software for AgiBot X1, a modular high-DOF humanoid robot. It runs neural-network locomotion policies via ONNX Runtime, bridges hardware drivers, and provides a simulation environment, all built atop AgiBot’s own AimRT middleware. In short, it is the C++ stack that translates reinforcement learning into joint commands, whether in a ROS2-based simulator or on the physical machine.

The interesting bit

Most open-source robotics projects lean on ROS2 as the primary backbone; here ROS2 is just one dependency among many, with AimRT handling the middleware heavy lifting. The README is unusually candid about the code being research-grade and frequently changing, which is a refreshing admission from a commercial hardware vendor.

Key highlights

• Full-stack release: model inference, platform drivers, and simulation bundled together
• ONNX Runtime for policy inference, with real-time Linux kernel support for hardware deployment
• Dual-mode operation: run_sim.sh for simulation, run.sh for the physical robot
• Joystick control module included for teleoperation
• Licensed under Mulan PSL-2.0, a Chinese open-source license rarely seen in Western robotics repos

Caveats

• The authors explicitly warn this is research code subject to frequent changes and not fit for all purposes
• Real-robot debugging requires a realtime kernel patch, and the default dependency downloads are slow enough that the repo ships a Gitee mirror fallback script
• Heavy toolchain: GCC-13, CMake ≥3.26, ROS2 Humble, and building ONNX Runtime from source are all expected

Verdict

Worth a look if you are building humanoid locomotion stacks or evaluating AimRT as middleware. Skip it if you wanted a lightweight, hardware-agnostic library—this is tightly coupled to the X1 and demands a serious build environment.