The humanoid for the data you don’t have yet.
Embodied-AI models don’t exist without embodied data. G1-D is Unitree’s purpose-built platform for collecting that data at scale — teleoperate, demonstrate, record, replay. Two 7-DoF arms, three HD cameras, on-device compute, and a battery that survives the shift.
Dual 7-DoF arms. Adaptive grippers. Interchangeable hands.
Two 7-degree-of-freedom arms with proprietary actuators and precision encoders, rated for ~3 kg per arm. Swap in a two-finger gripper for pick-and-place, a three-finger dexterous hand for delicate manipulation, or a full five-finger hand with tactile sensing for research-grade contact demos.
HD binocular head plus two wrist cameras.
One HD binocular camera at the head for scene context and stereo depth, plus two HD cameras at the wrists for close-up manipulation coverage. Wrist mounts see what the gripper is doing — critical for imitation learning where the training data has to include hand‑object interaction from the arm’s own point of view.
NVIDIA Jetson Orin NX. 100 TOPS. On-device inference.
A Jetson Orin NX 16GB module delivers 100 TOPS of on-device AI compute — enough to run vision transformers, imitation-learning policies and voice models locally without cloud round-trips. An 8-core high-performance CPU handles the robotic-control stack. Wi-Fi 6, Bluetooth 5.2 and API access ready.
Standard upper-body. Flagship mobile chassis.
Pick the format that fits the lab. The Standard upper-body is a benchtop workstation for stationary data-collection cells. The Flagship mounts the same upper body on a differential-drive base with LIDAR, dual depth cameras and 30 Ah of on-board battery — six hours of untethered mobile-manipulation demos, up to 1.5 m/s.
