[Paper Review] FlingBot: The Unreasonable Effectiveness of Dynamic Manipulation for Cloth Unfolding

Huy Ha, Shuran Song|arXiv (Cornell University)|May 8, 2021

Robot Manipulation and Learning25 references41 citations

TL;DR

FlingBot demonstrates that high-velocity dynamic flinging with a dual-arm pick-stretch-fling primitive can unfold cloth more efficiently than quasi-static methods, enabling large or unseen cloth types and real-world transfer.

ABSTRACT

High-velocity dynamic actions (e.g., fling or throw) play a crucial role in our everyday interaction with deformable objects by improving our efficiency and effectively expanding our physical reach range. Yet, most prior works have tackled cloth manipulation using exclusively single-arm quasi-static actions, which requires a large number of interactions for challenging initial cloth configurations and strictly limits the maximum cloth size by the robot's reach range. In this work, we demonstrate the effectiveness of dynamic flinging actions for cloth unfolding with our proposed self-supervised learning framework, FlingBot. Our approach learns how to unfold a piece of fabric from arbitrary initial configurations using a pick, stretch, and fling primitive for a dual-arm setup from visual observations. The final system achieves over 80% coverage within 3 actions on novel cloths, can unfold cloths larger than the system's reach range, and generalizes to T-shirts despite being trained on only rectangular cloths. We also finetuned FlingBot on a real-world dual-arm robot platform, where it increased the cloth coverage over 4 times more than the quasi-static baseline did. The simplicity of FlingBot combined with its superior performance over quasi-static baselines demonstrates the effectiveness of dynamic actions for deformable object manipulation.

Motivation & Objective

• Motivate the use of high-velocity dynamic actions to unfold deformable cloth more efficiently and beyond the robot's reach.
• Develop a self-supervised learning framework (FlingBot) that learns where to grasp for a two-arm fling to maximize cloth coverage.
• Demonstrate generalization to cloth types and sizes beyond training data, including cloths larger than arm reach and T-shirts.
• Fine-tune and validate on real dual-arm robots to compare against quasi-static baselines.

Proposed method

• Propose a two-arm pick, stretch, and fling primitive for cloth unfolding.
• Parameterize the dual-arm grasp with a compact four-scalar representation (center C, angle theta, width w) to satisfy reachability and collision constraints.
• Use spatial action maps to predict per-pixel value maps for grasp locations, invariant to translation, rotation, and scale.
• Train the value network in self-supervised simulation by regressing delta-coverage (change in cloth coverage) after actions.
• Fine-tune the trained policy on a real dual-arm UR5 setup and evaluate cloth coverage improvements.
• Provide a custom PyFlex-based simulator for cloth loading and a top-down RGB observation pipeline.

Experimental results

Research questions

• RQ1 Can dynamic fling actions with a dual-arm setup unfold severely crumpled cloth more efficiently than quasi-static methods?
• RQ2 Do two-arm fling grasps generalize across cloth types (rectangular, shirts) and reach beyond the robot's original workspace?
• RQ3 Does self-supervised learning from visual input suffice to learn effective fling grasp points without expert demonstrations?
• RQ4 How does real-world fine-tuning affect performance compared to simulation-only training?

Key findings

• FlingBot achieves over 80% cloth coverage within 3 actions on novel cloths in simulation.
• In simulation on Normal Rect cloths, FlingBot outperforms quasi-static baselines, increasing final coverage by 63.1 percentage points versus 24–29 points for baselines (approximate from table data).
• FlingBot maintains high performance on Large Rect and Shirt cloths, showing strong generalization to unseen geometries.
• In real-world tests, FlingBot reaches over 80% coverage across cloth types and improves over 40% relative to quasi-static pick & place baselines after finetuning.
• FlingBot significantly expands effective reach by enabling cloth unfolding beyond the robot’s initial reach range.
• The authors provide an open-source simulator and guidance for reproducing experiments.