A Review of Embodied Grasping

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Pre-trained models trained on internet-scale data significantly advance AI capabilities in perception, interaction, and reasoning.
• These models are increasingly foundational for embodied grasping methods, driving progress in robotics applications.

Purpose of the Study:

• To provide a comprehensive review of recent advancements in embodied grasping using pre-trained models.
• To summarize embodied foundations, algorithms, and future research directions in this field.

Main Methods:

• Review of cutting-edge embodied robots, simulation platforms, datasets, and data acquisition techniques.
• Analysis of embodied algorithms, focusing on pre-trained models for perception, strategy enhancement (imitation and reinforcement learning), and agent control.
• Exploration of how pre-trained models are used for point cloud extraction, 3D reconstruction, and direct action prediction.

Main Results:

• Pre-trained models enhance embodied perception by enabling better object understanding and environment awareness from sensor data.
• These models improve embodied strategy by enhancing data, acting as feature extractors in imitation learning, and optimizing reward functions in reinforcement learning.
• Pre-trained models facilitate embodied agents through hierarchical or holistic execution for end-to-end robot control.

Conclusions:

• The integration of pre-trained models offers significant potential for advancing embodied grasping and robot control.
• Addressing current research challenges and exploring feasible technical routes is crucial for future development in embodied AI.