Imitation learning for legged robot locomotion: a survey.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Imitation learning (IL) has transformed legged robot locomotion, eliminating the need for manual reward function engineering.
• Research in IL for locomotion has advanced significantly since 2019, moving from basic motion replication to complex policy generation with diffusion models.

Purpose of the Study:

• To provide a comprehensive survey of 35 key research works in imitation learning for legged robot locomotion.
• To analyze advancements in quadrupedal and humanoid robot platforms using a structured six-dimensional framework.
• To identify deployment challenges and future research directions in the field.

Main Methods:

• Systematic review and analysis of 35 research papers on imitation learning for legged robots.
• Categorization of studies based on robot platform (quadrupedal, humanoid) and methodology.
• Evaluation using a structured six-dimensional framework to assess progress and identify trends.

Main Results:

• Behavior cloning is employed in nearly half of the surveyed studies.
• Data generated via model-predictive control (MPC) is the most frequent training data source for advanced IL systems.
• Significant progress has been observed in policy generation, including the use of diffusion models.

Conclusions:

• Imitation learning is a dominant paradigm in legged robot locomotion research.
• Model-predictive control (MPC) is a key enabler for current advanced imitation learning systems.
• Further research is needed to address deployment challenges and explore novel research avenues.