Data-Driven Robotic Manipulation of Cloth-like Deformable Objects: The Present, Challenges and Future Prospects.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Science

Background:

• Cloth-like deformable objects (CDOs) present significant manipulation challenges in robotics.
• Their high degrees of freedom (DoF) lead to complex state-action dynamics and self-occlusion.
• Existing robotic control methods like imitation learning (IL) and reinforcement learning (RL) struggle with CDOs.

Purpose of the Study:

• To review data-driven control methods applied to CDO manipulation.
• To focus on four key task families: cloth shaping, knot tying/untying, dressing, and bag manipulation.
• To identify domain-specific inductive biases hindering generalizable IL and RL algorithms.

Main Methods:

• Literature review of data-driven control techniques for CDO manipulation.
• Analysis of application details across four major task families.
• Identification of challenges and inductive biases specific to each task domain.

Main Results:

• Data-driven methods show promise for CDO manipulation tasks.
• Significant challenges remain due to CDOs' inherent complexity.
• Specific inductive biases in cloth shaping, knot tying, dressing, and bag manipulation were identified.

Conclusions:

• Understanding domain-specific inductive biases is crucial for advancing robotic manipulation of CDOs.
• Further research is needed to develop more generalizable IL and RL algorithms for deformable objects.
• This review provides insights into current data-driven approaches and future research directions.