Dual-Level Humanoid Whole-Body Controller
This paper introduces Jaeger, a dual-level whole-body controller for humanoid robots that decouples upper and lower body control to overcome dimensionality challenges, utilizing a hybrid supervised and reinforcement learning approach trained on retargeted human motion data to achieve robust and versatile performance in both simulation and real-world environments.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Humanoid robots have long been the dream of science fiction, machines built to look and move like us, capable of navigating our stairs, opening our doors, and handling our tools. For these robots to be truly useful, they must master whole-body control, a complex balancing act where the machine coordinates dozens of joints simultaneously to stay upright while reaching, walking, or carrying objects. The challenge lies in the sheer number of moving parts; a robot's arms and legs often have conflicting needs, such as the arms swinging wildly to grab a cup while the legs must remain perfectly steady to prevent a fall. Traditional methods often struggle to teach a single computer program to manage all these moving pieces at once, leading to movements that are stiff, unstable, or prone to failure when the robot encounters real-world unpredictability.
A team of researchers at the Beijing Academy of Artificial Intelligence has developed a new approach to this problem, creating a control system they call JAEGER. Instead of forcing a single brain to manage the entire robot, they split the task into two distinct roles, much like how a human brain might separately manage the rhythm of walking while focusing the hands on a delicate task. This dual-level system assigns one controller to the robot's lower body, responsible for balance and movement across the ground, and a second, independent controller to the upper body, dedicated to reaching and manipulating objects. By separating these duties, the researchers found they could train the robot more effectively, allowing each part to focus on its specific job without getting in the way of the other.
The researchers tested this system on two different humanoid robots, the H1 and the H1-2, running thousands of training sessions in a simulated environment before moving to the real world. They taught the robots to perform a wide variety of tasks, from walking sideways and forward to executing complex dance-like movements that require precise coordination between the torso and limbs. In one demonstration, the robot used a joystick and virtual reality interface to approach and grab a doll while in motion, relying on the lower-body controller to maintain stability while the upper-body controller guided the arms. In another, the robot performed a series of dance moves, shifting its weight and moving its limbs in a fluid sequence that required both controllers to work in harmony.
To make this work, the team had to solve a difficult data problem: how to translate human movements, recorded from real people, into instructions a robot can follow. Human bodies and robot bodies are built differently; a human has a different number of joints and a different range of motion than a machine. The researchers created a fast, efficient system to map human poses onto the robot's skeleton, ensuring the movements looked natural and were physically possible for the machine. They then used a step-by-step training method, first teaching the robot to walk and balance on its own, and then gradually introducing the upper-body tasks. This approach prevented the robot from becoming confused or unstable during the learning process, allowing it to converge on a solution that was both robust and versatile.
The results of their experiments showed that this split-control strategy worked significantly better than previous methods. When tested on how well the robots could follow commands to move at specific speeds or hold specific poses, the new system made fewer errors than the leading alternatives. For instance, when asked to track a target speed while moving, the robot using the new system made errors of only about 0.15 meters per second in its forward speed, a marked improvement over other methods that struggled to stay on course. Similarly, when asked to mimic complex poses, the robot's joints moved with much greater precision, reducing the difference between the intended position and the actual position to a fraction of what other systems achieved.
The researchers also discovered that keeping the two controllers separate was crucial for the robot's stability. When they tried to use a single controller to manage the entire body, the robot often sacrificed balance for movement, or vice versa, leading to jerky or failed attempts at complex tasks. By decoupling the systems, the lower body could continue to function effectively even if the upper body encountered a problem, and the upper body could focus entirely on the task at hand without worrying about the mechanics of walking. This design also made the training process faster and more reliable, as the system could learn from human motion data without getting bogged down by the conflicting demands of a single, monolithic control network.
In the final tests, the robots demonstrated a remarkable ability to recover from disturbances. When researchers applied lateral forces equivalent to a push of up to 210 newtons, the robots were able to maintain their balance or recover from the stumble in more than 90 percent of the trials. This resilience suggests that the system is not just a theoretical success but a practical solution for real-world applications. The researchers noted that their method allows for high-frequency updates, meaning the robot can adjust its movements over a thousand times per second, a speed necessary for low-latency tasks like remote surgery or delicate manipulation.
The work does not claim to have solved every problem in robotics, but it offers a clear path forward for making humanoid robots more capable and reliable. By breaking down the complex task of whole-body control into manageable, specialized parts, the researchers have shown that robots can learn to move with a level of fluidity and adaptability that was previously difficult to achieve. The success of JAEGER suggests that the future of humanoid robotics may lie not in building a single, all-knowing brain, but in creating systems where different parts of the machine can work together, each doing what it does best, to achieve a common goal.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.