[Paper Review] Learning Robust Bed Making using Deep Imitation Learning with DART

Michael Laskey, Chris Powers+3 more|arXiv (Cornell University)|Nov 7, 2017

Robot Manipulation and LearningEngineering14 references17 citations

TL;DR

This paper proposes a robust bed-making system for mobile robots using deep imitation learning with DART, a method that enhances policy generalization by simulating robot errors during human demonstration collection. Trained on 50 demonstrations with DART, the robot achieves 96% sheet coverage even with distractors, outperforming a corner-detection baseline by 200%.

ABSTRACT

Bed-making is a universal home task that can be challenging for senior citizens due to reaching motions. Automating bed-making has multiple technical challenges such as perception in an unstructured environments, deformable object manipulation, obstacle avoidance and sequential decision making. We explore how DART, an LfD algorithm for learning robust policies, can be applied to automating bed making without fiducial markers with a Toyota Human Support Robot (HSR). By gathering human demonstrations for grasping the sheet and failure detection, we can learn deep neural network policies that leverage pre-trained YOLO features to automate the task. Experiments with a 1/2 scale twin bed and distractors placed on the bed, suggest policies learned on 50 demonstrations with DART achieve 96% sheet coverage, which is over 200% better than a corner detector baseline using contour detection.

Motivation & Objective

• Address the challenge of automating bed-making for senior citizens, who find the task physically demanding.
• Overcome technical hurdles in unstructured environments, including deformable object manipulation, obstacle avoidance, and sequential decision-making.
• Improve policy robustness to real-world perturbations such as household objects on the bed and lighting variations.
• Reduce reliance on large demonstration datasets by leveraging transfer learning from pre-trained YOLO features.
• Mitigate covariate shift in imitation learning by using DART to simulate and train recovery from small errors during data collection.

Proposed method

• Use a Toyota HSR robot to perform bed-making by grasping and pulling sheet corners toward the bed frame.
• Train deep neural network policies for grasp selection and transition success detection using human demonstrations.
• Leverage pre-trained YOLO features for visual state representation to improve sample efficiency.
• Apply DART (Dataset Aggregation with Robust Training) to inject small, optimized noise into human demonstrations, simulating robot errors during data collection.
• Use off-policy imitation learning to train policies that generalize to unseen states and recover from deviations.
• Integrate the robot’s internal motion planner for body movements, focusing imitation learning on vision-based decision-making components.

Experimental results

Research questions

• RQ1 Can a robot learn to make a bed robustly in the presence of common household distractors without fiducial markers?
• RQ2 How does DART improve policy generalization compared to standard behavior cloning in imitation learning for deformable object manipulation?
• RQ3 To what extent can transfer learning from pre-trained YOLO features reduce the number of required demonstrations for bed-making?
• RQ4 How does injecting simulated errors during demonstration collection reduce covariate shift and improve real-world performance?
• RQ5 What is the performance gain of DART-trained policies over heuristic methods like contour-based corner detection under distractor conditions?

Key findings

• DART-trained policies achieved 96% sheet coverage on a ½-scale bed with distractors (e.g., stuffed animals, toys), representing a 200% improvement over a corner-detection baseline using contour detection.
• With no distractors, DART-trained policies matched the supervisor’s performance within 2% in sheet coverage, demonstrating high fidelity imitation.
• DART reduced policy error on the robot’s actual execution distribution by exposing the supervisor to 27 failure modes during data collection, compared to only 2 in standard behavior cloning.
• The grasping policy trained with DART generalized better to unseen states due to exposure to a wider diversity of visual states during training.
• The transition policy showed significantly lower error on the robot’s distribution when trained with DART, indicating reduced covariate shift compared to behavior cloning.
• Despite requiring 30 seconds longer per demonstration on average, DART maintained 98% sheet coverage during data collection due to small, recoverable errors.