Unitree's G1 robot masterfully delivers kung-fu kicks with flawless balance VIDEO

Proceeding from its fresh article, Interesting Engineering highlights that a Chinese robotics company showcases its compact humanoid robot, G1, effortlessly executing kung-fu techniques. 

The video, released last week, reveals the robot’s newfound agility, precision, and control, thanks to a recent algorithm upgrade.

Unitree, the Chinese robotics company, has been consistently enhancing G1’s capabilities, enabling it to perform a wide variety of complex tasks. The latest update focuses on improving the robot's balance and expanding its movement range. G1's smooth dances, precise footwork, and ability to maintain stability even when interrupted highlight its enhanced agility.

Earlier this year, Unitree also demonstrated G1’s advanced walking and running system, which offers exceptional stability and control on uneven terrain and slopes. The robot's movement is increasingly fluid, showcasing how it can adapt to real-world environments.

The video of G1’s kung-fu performance showcases not only its martial arts skills but also its 23 degrees of freedom (DoF), allowing for remarkable coordination and flexibility. Despite its performance, the video concludes with a cautionary message: users should avoid modifying the robot in ways that could endanger others or teach it real combat skills.

Unitree positions the G1 as a versatile robot capable of performing demanding, repetitive tasks across various industries, including homes, factories, and hospitals. The company envisions humanoid robots as useful companions, supporting both work and everyday life.

Training for the G1 and other humanoid robots begins in a virtual environment with Nvidia’s Isaac Simulator. This method allows the robots to learn complex behaviors before they are physically constructed. The robots are trained by creating a digital twin that mimics human actions using motion capture and video data.

Through reinforcement learning, the robots practice and refine these behaviors in the virtual world. A technique called Sim2Real is then employed to transfer the learned skills into real-world applications, seamlessly integrating the virtual and physical realms.

In another effort to enhance its robots' natural movements, Unitree has released an open-source full-body dataset. This dataset, which is compatible with G1, H1, and H1-2 models, enables the robots to perform human-like motions, such as dancing, by improving their flexibility and coordination.

Captured with LAFAN1 motion capture technology, the dataset includes a redirection algorithm that optimizes the robots’ movements. The algorithm integrates interactive mesh and inverse kinematics technology, taking into account end posture constraints, joint positions, and velocity limitations to achieve fluid, realistic movement.

As Unitree continues to advance its humanoid robots, the G1 is poised to become a key player in robotics, offering both advanced capabilities and practical applications in various fields.

By Naila Huseynova