Advances in Humanoid Robotics and Manipulation

The field of humanoid robotics and manipulation is rapidly evolving, with a focus on developing more efficient, robust, and adaptable control policies. Recent research has emphasized the importance of integrating learning-based methods with model-based approaches to reduce training complexity and ensure safety and stability. Notable developments include the use of adversarial policy learning, physics-informed world models, and sparse anchor posture curriculum learning to improve whole-body control, non-prehensile manipulation, and motion synthesis. These innovations have the potential to significantly advance the field and enable more sophisticated humanoid robot capabilities. Noteworthy papers include: Adversarial Locomotion and Motion Imitation for Humanoid Policy Learning, which proposes a novel framework for adversarial policy learning between upper and lower body. PIN-WM: Learning Physics-INformed World Models for Non-Prehensile Manipulation, which presents a physics-informed world model for robust policy learning and generalization. ProMoGen: Progressive Motion Generation via Sparse Anchor Postures Curriculum Learning, which introduces a novel framework for integrating trajectory guidance with sparse anchor motion control.

Sources

Adversarial Locomotion and Motion Imitation for Humanoid Policy Learning

Accelerating Visual Reinforcement Learning with Separate Primitive Policy for Peg-in-Hole Tasks

PIN-WM: Learning Physics-INformed World Models for Non-Prehensile Manipulation

PMG: Progressive Motion Generation via Sparse Anchor Postures Curriculum Learning

Demonstrating Berkeley Humanoid Lite: An Open-source, Accessible, and Customizable 3D-printed Humanoid Robot

Integrating Learning-Based Manipulation and Physics-Based Locomotion for Whole-Body Badminton Robot Control

Built with on top of