[Paper Review] Visual Imitation Made Easy

Sarah Young, Dhiraj Gandhi+4 more|arXiv (Cornell University)|Aug 11, 2020

Robot Manipulation and LearningEngineering42 references22 citations

TL;DR

This paper proposes a low-cost, scalable interface for visual imitation learning using commercially available reacher-grabber tools to collect diverse, real-world demonstrations. By leveraging Structure from Motion (SfM) and a finger detection network to extract 3D tool trajectories and gripper states, the method enables direct behavior cloning with data augmentation, achieving 87.5% success on pushing and 62.5% on stacking with unseen objects—demonstrating that large-scale, diverse data collection with simple tools can significantly improve policy generalization.

ABSTRACT

Visual imitation learning provides a framework for learning complex manipulation behaviors by leveraging human demonstrations. However, current interfaces for imitation such as kinesthetic teaching or teleoperation prohibitively restrict our ability to efficiently collect large-scale data in the wild. Obtaining such diverse demonstration data is paramount for the generalization of learned skills to novel scenarios. In this work, we present an alternate interface for imitation that simplifies the data collection process while allowing for easy transfer to robots. We use commercially available reacher-grabber assistive tools both as a data collection device and as the robot's end-effector. To extract action information from these visual demonstrations, we use off-the-shelf Structure from Motion (SfM) techniques in addition to training a finger detection network. We experimentally evaluate on two challenging tasks: non-prehensile pushing and prehensile stacking, with 1000 diverse demonstrations for each task. For both tasks, we use standard behavior cloning to learn executable policies from the previously collected offline demonstrations. To improve learning performance, we employ a variety of data augmentations and provide an extensive analysis of its effects. Finally, we demonstrate the utility of our interface by evaluating on real robotic scenarios with previously unseen objects and achieve a 87% success rate on pushing and a 62% success rate on stacking. Robot videos are available at https://dhiraj100892.github.io/Visual-Imitation-Made-Easy.

Motivation & Objective

• To address the bottleneck in large-scale, diverse data collection for visual imitation learning by replacing restrictive interfaces like teleoperation or kinesthetic teaching.
• To enable efficient, real-world data collection using low-cost, widely available assistive tools such as reacher-grabbers.
• To demonstrate that direct behavior cloning from visual demonstrations collected with these tools can generalize to novel objects and environments.
• To evaluate the impact of data augmentation and data diversity on policy performance in real robotic scenarios.

Proposed method

• Use commercially available reacher-grabber tools as both the data collection interface and the robot’s end-effector to ensure direct transferability.
• Collect RGB video demonstrations from a camera mounted on the reacher-grabber during task execution in diverse home and office environments.
• Extract 3D tool trajectories using off-the-shelf Structure from Motion (SfM) techniques from the collected video sequences.
• Detect gripper finger states using a trained finger detection network to infer action-level information from visual demonstrations.
• Train policies via standard behavior cloning using the extracted trajectories and actions, with data augmentation to improve generalization.
• Transfer the learned policies directly to a robot with the same end-effector setup, enabling zero-shot deployment on unseen objects.

Experimental results

Research questions

• RQ1 Can assistive tools like reacher-grabbers be used as a scalable, low-cost interface for collecting diverse visual demonstrations for robot learning?
• RQ2 How does data diversity and scale from real-world demonstrations affect policy generalization in visual imitation learning?
• RQ3 To what extent do data augmentation techniques improve policy performance on unseen objects in non-prehensile pushing and prehensile stacking tasks?
• RQ4 Can behavior cloning from visual demonstrations collected with assistive tools outperform traditional methods in real-world robotic deployment?

Key findings

• The proposed DemoAT framework achieved 87.5% success rate on non-prehensile pushing and 62.5% on prehensile stacking with previously unseen objects, significantly outperforming naive behavior cloning.
• Data augmentation, particularly crop + jitter for pushing and rotation + jitter for stacking, improved performance by 0.7% and 0.9% on the BC-MSE metric, respectively.
• Using data augmentations with only 50% of the training data surpassed the performance of naive behavior cloning with 100% of the data on both pushing and stacking tasks.
• The model trained on 50% of the data with augmentations achieved 100% success in reaching the first object in the pushing task, demonstrating strong generalization from limited data.
• A diverse dataset with sparse observations across many scenes outperformed a less diverse dataset with repeated scenes, with a 1.4% lower test error (0.067 vs. 0.081).
• Performance gains from data augmentation were most pronounced in the later stages of tasks, with a 33.4% improvement in stacking success rate after the initial reach phase.