Humanoid Locomotion and Manipulation: Current Progress and Challenges

Abstract

Comprehensive survey covering the state-of-the-art in humanoid locomotion and manipulation, with focus on control, planning, and learning methods. Examines advances across model-based methods, learning approaches, and emerging technologies, addressing trade-offs between model fidelity and computational efficiency.

Key Contributions

• Comprehensive methodology overview spanning three decades of model-based humanoid robotics
• Integrated perspective on combined locomotion-manipulation capabilities
• Learning paradigm analysis including reinforcement and imitation learning
• Foundation models integration potential for generalist agents
• Tactile sensing emphasis as critical for contact-rich interactions

Methodology

Covers: contact planning, motion planning, whole-body control, reinforcement learning, imitation learning, foundation model integration, and tactile sensing/feedback.

Results

Rapid progress driven by advances in machine learning and existing model-based approaches, though typically developed separately rather than in unified frameworks. Identifies trend toward learning-based methods but notes model-based approaches remain essential for safety-critical scenarios.

Limitations

• Balancing robustness across different operational scenarios
• Computational efficiency constraints
• Versatility and generalization across varied tasks
• Integration of disparate control and learning paradigms
• Effective utilization of whole-body tactile feedback

Source: Humanoid Locomotion and Manipulation by Gu et al.