A review of learning-based dynamics models for robotic manipulation.

Area of Science:

• Robotics
• Machine Learning
• Physics Simulation

Background:

• Accurate dynamics models are crucial for robot planning and control.
• Physics-based models often require full-state information, which is difficult to obtain in real-world scenarios.
• Learned dynamics models offer an alternative by learning from perceived interaction data.

Purpose of the Study:

• To provide a comprehensive review of current techniques for learned dynamics models in robotics.
• To highlight the trade-offs in designing these models.
• To identify research gaps and future directions.

Main Methods:

• Review of existing literature on learning-based dynamics models for robotics.
• Analysis of state representation techniques and their impact on inductive biases.
• Discussion of integration with state estimation and control.

Main Results:

• Learned dynamics models capture complex factors, predictive uncertainty, and accelerate simulations.
• Advancements in robot capabilities for manipulating deformable objects, granular materials, and multi-object interactions.
• The choice of state representation is critical for reduced-order modeling.

Conclusions:

• Learned dynamics models are advancing robot manipulation capabilities.
• Integration with state estimation and control is key for practical applications.
• Further research is needed to address critical gaps in the field.