From the perspective of application:

Humanoid robots are built to (ideally) replicate human motions in performing various human-level skills, such as locomotion, manipulation, and cognitive capabilities.

From the perspective of composition:

A humanoid robot refers to any anthropomorphic robot that resembles the form of a human. Typically, a humanoid robot possesses a torso, two arms, and two legs, though the degree of anthropomorphism may vary.

Background

Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning lists background, i.e., definition of locomotion, manipulation, loco-manipulation, whole-body control.

Learning Loco-manipulation Skills

Reinforcement Learning from Scratch

Refer to Reinforcement Learning for Humanoid Robots.

Learning from Demonstration

Refer to Imitation Learning for Humanoid Robots.

Hybrid Models

  • A teacher policy trained from simulation using pure RL. Then a student policy clones the behavior of the teacher.

  • Using IL first to pre-train an policy from demonstration. Then a RL policy fine-tunes the policy.

Representaion of Skills

  • motion Representation

  • goal representation

  • state transition representation

Learning for Humanoid Loco-manipulation

  • most in simulation, the physical interactions with external environment or objects are often oversimplified

  • careful reward design using RL

Foundation Models for Humanoid Robots

Applying LLMs/VLMs to Humanoid Robots

Refer to Real-World Robot Applications of Foundation Models: A Review.

Building Humanoid Foundation Models

Refer to Robotic Foundation Models for robotic foundation models.