G1

The Unitree G1 humanoid robot, developed by Unitree Robotics, is a compact, versatile platform standing 132 cm tall (folded to 69 cm) and weighing approximately 35 kg, designed for advanced mobility and manipulation. Its architecture features 23 degrees of freedom (DOF) in the standard model, expandable to 43 in the EDU version, including 6 DOF per leg, 1-3 DOF at the waist, 5 DOF per arm, and up to 9 DOF per hand with optional dexterous grippers like Dex3-1 (three-fingered with tactile sensors). Joints utilize low-inertia, high-speed permanent magnet synchronous motors (PMSM) with maximum knee torque of 120 N.m (EDU), dual encoders for precise position/velocity feedback, industrial-grade crossed roller bearings, full hollow electrical routing, and local air cooling for thermal management. Sensing is powered by a Livox MID-360 3D LiDAR (360° horizontal, 59° vertical FOV), Intel RealSense D435i depth camera, 4-microphone array, and 5W stereo speakers, enabling perception, voice interaction, and navigation. Compute includes an 8-core high-performance CPU standard, with NVIDIA Jetson Orin NX (8-core Arm Cortex-A78AE @2GHz, 16GB RAM, 1024 CUDA cores, 2TB storage) optional in EDU for secondary development. Power comes from a quick-release 9000 mAh, 54V (13S lithium-ion) smart battery, providing ~2 hours runtime, charged via 54V 5A adapter. Connectivity: WiFi 6, Bluetooth 5.2, Gigabit Ethernet. AI architecture centers on UnifoLM (Unified Robot Large Model), a proprietary world model integrating imitation learning, reinforcement learning (RL), and force-position hybrid control for dexterous manipulation simulating human hands. It supports OTA updates, SLAM/navigation via LiDAR/depth fusion, and SDK for ROS2/DDS interfaces, motion services, RL control, and high-level actions like arm control. No native LLM integration, but extensible for custom embodied AI. Real-world deployments include viral demos: world's first kip-up and side-flip (2025), martial arts/boxing tournaments with multiple G1s, live concert dances, CES 2025 hands-on, ETH Zurich behavior imitation training, NVIDIA Isaac Lab rough-terrain locomotion sim-to-real, CMU SPARK dexterity tasks, DoorMan door-opening outperforming teleop. Research highlights security vulnerabilities (e.g., Alias Robotics espionage risks) and sim benchmarks. Primarily research/education-focused, with emerging industrial potential in warehouses via G1-D variant. Continuous evolution via OTA positions G1 as an accessible humanoid for AI-robotics advancement.