Awesome-Robotics-Manipulation

• [2025/08] Major revision of the classification system with a more refined taxonomy; substantial improvements across all sections.
• [2025/07] Expanded coverage of Dexterous, Soft Robotic, Mobile, Quadrupedal, and Humanoid Manipulation; refined the categorization and content for Awesome Simulators, Benchmarks, and Datasets; added non-learning-based control methods.
• [2025/06] Introduced new sections on Grasp in Cluttered Scenes, Quadrupedal and Humanoid Manipulation, and Learning from Human Demonstrations. Also improved the classification of the Applications section and added a subsection on Embodied QA Datasets.
• [2025/02] Added a new section on Bimanual Grasp.
• [2024/12] Introduced coverage of Dexterous Grasp.
• [2024/10] Repository is now public!

This survey presents a unified perspective on robot manipulation by organizing existing methods according to the relationship between high-level planning and low-level action modeling. We provide a systematic taxonomy that connects different forms of task-level reasoning, structured representations, and executable action generation.

Summary. The survey first reviews robot manipulation from multiple aspects, including task types, robot embodiments, simulators and benchmarks, high-level planning, learning-based action modeling, applications, and key challenges. This overview provides a comprehensive map of recent advances and highlights the connections among different research directions.

At the algorithmic level, we formulate robot manipulation as a two-stage process consisting of high-level planning and low-level action modeling. High-level planning focuses on generating structured intermediate representations, such as task plans, geometric constraints, affordances, code, video predictions, and 3D representations. These structured outputs can be instantiated as constraints or inputs for low-level action models, which generate executable robot actions.

For learning-based action modeling, we further organize existing methods into three fundamental components: input modeling, which studies what sensory modalities are used and how they are represented; latent learning, which explores how intermediate representations or latent actions are learned; and policy learning, which investigates how actions are generated from learned representations.

Literature Distribution. The survey covers a broad range of literature from major robotics, machine learning, and computer vision venues. The collected papers are mainly published in leading conferences and journals, including CoRL, ICRA, RSS, IROS, NeurIPS, ICLR, CVPR, ICML, IEEE RA-L, IEEE T-RO, IJRR, and other representative venues.

Word Cloud. The keyword analysis summarizes major research trends in robot manipulation, highlighting emerging topics such as imitation learning, reinforcement learning, vision-language-action models, diffusion policies, dexterous manipulation, grasping, world models, and generalization.