Robust Visuomotor Control for Humanoid Loco-Manipulation Using Hybrid Reinforcement Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Humanoid robots are valuable for practical tasks.
• Reinforcement learning (RL) offers robust control but faces challenges in high-dimensional, long-horizon loco-manipulation tasks.
• Visuomotor control in these tasks is particularly difficult.

Purpose of the Study:

• To propose a novel loco-manipulation control framework for humanoid robots.
• To address the challenges of high dimensionality and long-horizon exploration in RL-based visuomotor control.
• To improve the adaptability and success rate of humanoid robots in complex tasks.

Main Methods:

• Utilized model-free reinforcement learning (RL) integrated with model-based control in the robot's task space.
• Implemented a visuomotor policy using depth-image input.
• Employed mid-way initialization and prioritized experience sampling to accelerate policy convergence.

Main Results:

• Achieved an 83% overall success rate on typical loco-manipulation tasks, including load carrying and door opening.
• Demonstrated the framework's ability to automatically adjust robot motion in response to environmental changes.
• Validated the effectiveness of the proposed RL-based visuomotor control approach.

Conclusions:

• The proposed framework significantly enhances humanoid robot capabilities in loco-manipulation tasks.
• The integration of model-free RL with model-based control, coupled with specific optimization techniques, overcomes key challenges.
• The system's adaptability to environmental dynamics is crucial for real-world applications.