Unitree’s latest demonstration of its compact G1 humanoid robot showcases groundbreaking advancements in agility, intelligence, and adaptability. In a newly released video, the Chinese robotics firm highlights how its cutting-edge AI-driven algorithms have enabled G1 to perform complex movements with remarkable precision—signaling a new era for humanoid robotics.
The G1 has already made history by executing the world’s first kick-up, displaying unprecedented agility in a humanoid robot. Last week, Unitree demonstrated another milestone: the G1 successfully performed a standing side flip, surpassing the backflip achieved by its larger counterpart, the H1, in March 2024.
The latest video further emphasizes the G1’s advanced motor control, showcasing its ability to perform combat-style sweeping kicks, intricate Tai Chi moves, and a recovery maneuver from a face-up position despite visual obstructions. Additionally, the robot demonstrated impressive reflexes when pushed from behind, instinctively adjusting its limbs to maintain balance—mimicking human-like responses.
Unitree’s training approach involves a hybrid of simulation and real-world application. The G1 begins its learning process in Nvidia’s Isaac Simulator, where a digital twin of the robot is created using motion capture data. Through reinforcement learning, its movements are refined before being transferred into the physical robot via a technique known as Sim2Real.
On the hardware front, the G1 boasts 23 degrees of freedom, ensuring fluid movement and precise coordination. Standing at 4.33 feet (1.32 meters) and weighing 77 lbs (35 kg), it features a foldable design for easy storage. Equipped with 3D LiDAR, a RealSense depth camera, and a noise-canceling microphone array for voice commands, the robot is designed for real-world interaction. A high-performance CPU and a 9,000-mAh battery allow for up to two hours of operation, with fast battery-swapping capabilities enabling continuous use.
Unitree is also contributing to the broader robotics community by releasing an open-source motion dataset. Developed using LAFAN1 motion capture technology and advanced redirection algorithms, this dataset enhances humanoid robots’ motion planning, improving their ability to perform dynamic tasks like dancing and repositioning.
