How to build QingLoong, OpenLoong's full-size open-source reference humanoid
16 steps to build QingLoong, OpenLoong's full-size open-source reference humanoid, with the source documents.
Not sold; the official pages give no price and no BOM estimate (2026‑11‑03)
Buy parts
- Select Spherical Navigation Radar
Select the Speedte Bpearl spherical radar as the navigation radar; its 360° detection angle covers most of the area in front of the robot and around its feet.
- Select Blind-Spot Depth Cameras
Select depth cameras with a 50 mm stereo baseline, 19 cm–10 m operating range, depth accuracy below 2% at 2 m, and H80.5°×V55.2° field of view. They provide 640×400 depth images at 30 fps over USB 2.0, use 5 VDC power, and consume 1 W average with less than 6 W peak power.
- Select Infrared Camera CMOS
Select a 720P infrared camera CMOS with H1280×V720 resolution, 3 μm×3 μm pixels, global exposure, and 30 fps. It uses 940 nm infrared with an adapted-lens field of view of H80.5°×V55.2°.
- Select Color Camera CMOS
Select a color camera CMOS with a 1280×720 pixel array and MIPI interface, using 1.75 μm×1.75 μm pixels. It uses rolling exposure at 30 fps, 650 nm infrared, and an adapted-lens field of view of H65.9°×V51.3°.
- Select Depth Camera Board
Select a board powered by USB 2.0 that runs Linux and supports three MIPI cameras, with depth-map performance of 30fps@640×400.
- Select Four Panoramic Modules
Select GC2035 camera modules for the four ultra-wide visible-light cameras installed around the waist.
- Download and Extract Package
Download and extract the project archive.
- Download and unzip the files
Download and unzip the project archive.
Assemble
- Mount Radar Ahead of Neck
Mount the navigation radar in front of the robot's neck to incrementally build maps of the area ahead and support autonomous navigation.
- Add Cameras at Front and Rear Waist
Add one depth camera at the front and one at the rear of the robot's waist to detect obstacles behind and below the robot that the chest lidar cannot perceive.
- Add Chest and Back Cameras
As needed, install one wide-angle camera at the front chest and one at the back, connecting them to the video processing board through SDI. They can directly or jointly output the forward and rear environmental views.
- Assemble Hardware from Drawings
Assemble the hardware according to the drawings and documents in the directory.
- Assemble the hardware from drawings
Assemble the hardware using the drawings and documents in the directory.
Software
- Integrate Control and Sensor Software
Integrate the software using the included control system and sensors.
- Integrate and debug control software
Use the accompanying control system and sensors for software integration and debugging.
Calibrate and test
- Debug Robot Hardware and Software
Complete debugging with the included control system and sensors.
Parts
- Peak joint torques (PDF table 6): hip pitch and knee 396 N·m, hip roll 320 N·m, hip yaw 160 N·m, ankle 208 N·m; waist pitch and roll 315 N·m, waist yaw 121 N·m; shoulder joints 1–2 80 N·m, shoulder 3 and elbow 4 58 N·m, elbow 5 and wrist 6–7 11.5 N·m; neck 26 N·m
- EtherCAT servo drives: 11–95 VDC, ≥130 A, current loop ≥10 kHz; the main controller reaches the joints through upper- and lower-body EtherCAT branch units
- A 72V (19S) battery of ≥30 Ah, ≤9 kg and ≥120 A maximum discharge, with a BMS; the wiring diagram adds 72V-to-48V and 24V-to-12V buck modules
- Motion-control computer: up to 400 TOPS, 4 gigabit Ethernet ports, Wi‑Fi 6, 4G with GPS/BeiDou, 8 CAN buses and at least 6 USB3 ports
- Humanoid dexterous hands: 19 DoF (6 active), ≤600 g each, 8.4V rated, with a quick-change mount
- Perception: a 3D lidar on the chest, a stereo camera and display in the head, blind-spot depth cameras at the front and back of the waist, and four GC2035 wide-angle surround cameras on the waist
Tools
- Outsourced machining from the PDF drawings and STEP files for the waist, chest, head-perception and leg-and-foot parts (the V2.5 STEP files need Git LFS)
- Wiring from the “QingLoong 2.1” electrical diagram: two emergency stops, a remote switch and a relay, with STO lines to the joint drives
- A Linux computer with the IGH EtherCAT master; Isaac Gym training needs a CUDA GPU (Gymloong recommends 16 GB of VRAM or more)
Skills
- Wiring and commissioning an EtherCAT bus and a 72V battery system
- MPC and WBC whole-body control (OpenLoong-Dyn-Control in MuJoCo)
- Reinforcement learning in Isaac Gym (Gymloong: Ubuntu 20.04, Python 3.8)
Software
- OpenLoong-Dyn-Control (Apache-2.0): whole-body dynamics control on MPC and WBC, with walking, jumping and blind obstacle-stepping demos in MuJoCo
- OpenLoong-Framework (Apache-2.0): a C++ control framework without ROS, with the same interface in simulation and on the robot, bundling every dependency except the IGH EtherCAT master
- OpenLoong-Gymloong (Apache-2.0): reinforcement learning tutorial and code for training QingLoong to walk in Isaac Gym