Humanoid robots from the Chinese company Unitree have made headlines many times – they can do backflips, dance, crack nuts and toast. At the end of February, the G1 model’s repertoire of movements expanded again, thanks to which the robot mastered kung fu techniques. Today, G1, with 23 degrees of freedom, set a new record – it demonstrated a flawless performance of a side somersault from a standing position.
Image source: Unitree
In March 2024, G1’s predecessor, the Unitree H1, wowed viewers by performing a backflip. Today, Unitree released a 19-second video demonstrating G1’s exceptional balance and precision while performing an even more challenging feat: a standing sideflip.
Unitree constantly updates the robot’s algorithms, allowing it to learn and master increasingly complex and precise movements. The G1’s design provides 23 degrees of freedom, which gives the robot impressive stability and coordination. It uses 3D lidar, a depth camera, and a noise-canceling microphone array to recognize voice commands. The heart of the robot is an 8-core processor, which provides high maneuverability.
The G1 is equipped with a quick-release 9000 mAh battery, which provides up to two hours of operation with the ability to quickly replace it. The maximum speed of the G1 is 7.2 kilometers per hour. The humanoid robot, with a height of 1.32 meters and a weight of 35 kg, gracefully folds for storage in a container measuring only 69 × 44 × 30 centimeters.
G1 is initially trained using Nvidia’s Isaac simulator. This allows the robot to learn highly complex behaviors in a controlled digital environment using reinforcement learning. The learned actions are then seamlessly transferred to a physical model using a process called Sim2Real, ensuring that movements are executed accurately in the real world.
To accelerate progress in humanoid robotics, Unitree has released an open-source dataset aimed at improving the control and coordination of humanoid robots. The dataset improves the smoothness of movements of Unitree’s core models, including the H1, H1-2, and G1, and facilitates “human-like” movements such as dancing and dynamic posture changes.
The dataset, developed using LAFAN1 motion capture technology, is fully compatible with Unitree humanoid systems and includes an advanced redirection algorithm. This algorithm optimizes motion planning using interactive processing and inverse kinematics, taking into account pose constraints, joint articulations, and velocity parameters.