FALCON: Learning Force-Adaptive Humanoid Loco-Manipulation (2505.06776v1)

Abstract: Humanoid loco-manipulation holds transformative potential for daily service and industrial tasks, yet achieving precise, robust whole-body control with 3D end-effector force interaction remains a major challenge. Prior approaches are often limited to lightweight tasks or quadrupedal/wheeled platforms. To overcome these limitations, we propose FALCON, a dual-agent reinforcement-learning-based framework for robust force-adaptive humanoid loco-manipulation. FALCON decomposes whole-body control into two specialized agents: (1) a lower-body agent ensuring stable locomotion under external force disturbances, and (2) an upper-body agent precisely tracking end-effector positions with implicit adaptive force compensation. These two agents are jointly trained in simulation with a force curriculum that progressively escalates the magnitude of external force exerted on the end effector while respecting torque limits. Experiments demonstrate that, compared to the baselines, FALCON achieves 2x more accurate upper-body joint tracking, while maintaining robust locomotion under force disturbances and achieving faster training convergence. Moreover, FALCON enables policy training without embodiment-specific reward or curriculum tuning. Using the same training setup, we obtain policies that are deployed across multiple humanoids, enabling forceful loco-manipulation tasks such as transporting payloads (0-20N force), cart-pulling (0-100N), and door-opening (0-40N) in the real world.

Overview of FALCON: Learning Force-Adaptive Humanoid Loco-Manipulation

The paper "FALCON: Learning Force-Adaptive Humanoid Loco-Manipulation" introduces an innovative framework for enhancing force-adaptive humanoid robotic tasks. Developed using a dual-agent reinforcement learning architecture, FALCON is designed to address significant challenges in achieving robust and precise whole-body control for humanoid robots engaged in dynamic loco-manipulation tasks. The framework focuses on decomposing the loco-manipulation process into lower-body and upper-body control, each governed by distinct policies and learning mechanisms to handle interaction forces efficiently.

Technical Contributions

FALCON's primary innovation lies in its architecture that deploys two cooperative agents trained under force-disturbance conditions:

• Lower-Body Agent: Responsible for maintaining stable locomotion amidst external perturbations. This agent ensures the robot can walk, squat, and twist while transporting payloads under varying forces.
• Upper-Body Agent: Focuses on precise manipulation by compensating for 3D external forces acting on the end-effector, thus enhancing joint tracking accuracy.

These agents are trained with a force curriculum that gradually escalates the external forces exerted to simulate real-world conditions, ensuring that the framework respects the torque limits of the humanoid structure.

Experimental Results

Quantitative evaluations demonstrate FALCON's superior performance compared to existing approaches, notably providing a twofold increase in upper-body joint tracking accuracy while sustaining robust locomotion. Additionally, the framework accomplishes faster training convergence, which is a significant advantage in the field of reinforcement learning for robotics. This efficiency is achieved without requiring embodiment-specific rewards or curriculum tuning, allowing policies trained under identical setups to be applicable across multiple humanoid platforms.

Implications and Future Directions

The applications of FALCON are extensive, particularly in tasks such as warehouse logistics and service industries where humanoids might perform various complex interactions involving significant forces. Practically, the framework promises improvements in scenarios requiring adaptable manipulation and transport of objects, such as cart-pulling and door-opening.

Theoretical implications involve advancements in understanding and modeling the dynamics of humanoid robots under forceful interactions, contributing to robotics research's ongoing quest for versatility and robustness in human-like tasks.

Future research might focus on extending force-adaptive strategies to accommodate torque disturbances not limited to end-effectors. Considering rotational force disturbances and expanding the framework to multi-contact interactions will potentially enhance the functionality and deployment of humanoid systems in even more varied environments.

In conclusion, FALCON presents a noteworthy advancement in the field of humanoid robotics, offering a robust framework for adaptive loco-manipulation. Its methodology and results pave the way for enhanced humanoid capabilities, contributing significantly to applied and theoretical robotics research.