HUGWBC - A humanoid robot full-body controller jointly developed by Shanghai Jiao Tong University and Shanghai AI Lab.

What is HUGWBC?

HUGWBC (Humanoid Unified and General Whole-Body Controller) is a humanoid robot whole-body controller jointly developed by Shanghai Jiao Tong University and Shanghai AI Lab, enabling precise motion control. Based on an expanded command space and advanced training techniques, HUGWBC allows robots to perform various natural gaits (such as walking, running, jumping, standing, and hopping on one leg), supports real-time external upper limb control signals, and achieves complex motor tasks. HUGWBC is trained using reinforcement learning in a simulated environment and uses an asymmetric training framework (AAC) to directly transfer policies to real robots.

HUGWBC's main functions

• Supports multiple gait modesIt supports controlling robots to perform various gaits such as walking, running, jumping, standing, and hopping on one leg.
• Fine-tuning parametersIt supports adjusting parameters such as gait frequency, foot swing height, body height, waist rotation, and body pitch angle to adapt to different task requirements.
• Real-time external interventionIt supports real-time external control signals for the upper limbs (such as teleoperation or motion redirection), enabling the robot to perform complex operations when executing motion tasks.
• High robustness and adaptabilityHUGWBC can adapt to different terrains and external disturbances, maintaining stability and accuracy in motion.

HUGWBC Technical Principles

• Extended command spaceDesign a general command space that includes task commands and behavior commands to flexibly control the robot's movement and posture.
• Reinforcement learning trainingThe control strategy is trained in a simulated environment based on reinforcement learning (especially the PPO algorithm), and a reward function is defined to optimize the robot's motion performance.
• Asymmetric training frameworkThe Asymmetric Training Framework (AAC) is adopted, which allows the critique network to access all privileged information, while the actor network uses only sensor data available to real robots, thereby improving the generalization ability of the policy.
• Symmetry lossIntroducing a symmetric loss function encourages the policy to generate natural and symmetrical motions, thereby improving the stability and efficiency of the motion.
• Intervention trainingIntroduce external intervention signals during training to adapt to real-time control of the upper limbs while maintaining the stability and precision of lower limb movements.
• Foot trajectory planningThe fifth-order polynomial trajectory planning method is used to ensure smooth foot movement during the swing phase, thereby improving the naturalness and stability of the movement.

HUGWBC project address

• Project official website:https://hugwbc.github.io/
• arXiv technical paper:https://arxiv.org/pdf/2502.03206

Application scenarios of HUGWBC

• Navigation in complex terrainControlling humanoid robots to walk, run, and jump on uneven terrain, adapting to various complex environments such as rugged mountain roads, ruins, or construction sites, and performing search and rescue missions.
• Dynamic task executionIt enables robots to perform highly dynamic tasks, such as running, jumping, and hopping on one leg, playing a role in scenarios requiring rapid response and high mobility, such as sports event assistance or emergency response tasks.
• Human-machine collaborative tasksBased on real-time external upper limb intervention, humanoid robots can collaborate with human operators to complete complex tasks, such as handling, assembly, or fine operations, thereby improving work efficiency and safety.
• Service robot applicationsIn homes or service settings, humanoid robots can perform everyday tasks such as cleaning, moving items, or assisting the elderly and disabled, providing more natural and flexible services.
• Scientific research and developmentTo develop and test new motion control algorithms, human-computer interaction technologies, and robot perception capabilities, thereby promoting the further development of humanoid robot technology.