Humanoid Robots Master Tug of War and Door Opening With New AI System
- tech360.tv
- May 20, 2025
- 2 min read
Humanoid robots are learning to walk, pull, and push with greater precision and balance, thanks to a new AI control system developed by Carnegie Mellon University.
The system, called FALCON, enables robots to perform complex tasks like pulling carts, opening doors, and carrying loads while maintaining stability under unpredictable forces.
Developed by researchers at CMU’s Learning and Control for Agile Robotics Lab, FALCON uses two separate AI agents—one for walking and one for arm control. These agents are trained together in simulation, allowing the robot to coordinate full-body movements with force-aware precision.
FALCON was tested on two humanoid robots, Unitree G1 and Booster T1, and showed strong performance across different platforms without the need for special tuning.
The system allows robots to handle end-effector forces up to 100 newtons, about 30% of the robot’s body weight, while maintaining balance and improving arm control accuracy by 100% compared to previous methods.
FALCON’s training method gradually increases the 3D force applied to the robot’s hands while ensuring joint safety through physics-based force checks. This enables robots to learn stable walking and precise arm movements even during forceful tasks.
The system’s dual-agent reinforcement learning framework addresses limitations in earlier approaches, which either lacked unified force control or suffered from slow, ineffective learning.
Despite its strengths, FALCON has limitations. It does not manage contact forces on body parts other than the hands, making it less effective for tasks like leaning or collaborative lifting. It also struggles with rotational disruptions, such as twisting handles or using tools, due to its focus on external forces rather than torques.
Researchers suggest future improvements should include torque-aware policies and multi-contact reasoning to expand FALCON’s capabilities.
The full research details are available on GitHub.
CMU developed FALCON to improve humanoid robot balance and force control
The system uses two AI agents for walking and arm movement
Robots trained with FALCON can pull carts, open doors, and carry loads
Source: INTERESTING ENGINEERING
