A Wearable Robotic Hand for Hand-over-Hand Imitation Learning

Abstract: Dexterous manipulation through imitation learning has gained significant attention in robotics research. The collection of high-quality expert data holds paramount importance when using imitation learning. The existing approaches for acquiring expert data commonly involve utilizing a data glove to capture hand motion information. However, this method suffers from limitations as the collected information cannot be directly mapped to the robotic hand due to discrepancies in their degrees of freedom or structures. Furthermore,it fails to accurately capture force feedback information between the hand and objects during the demonstration process. To overcome these challenges, this paper presents a novel solution in the form of a wearable dexterous hand, namely Hand-over-hand Imitation learning wearable RObotic Hand (HIRO Hand),which integrates expert data collection and enables the implementation of dexterous operations. This HIRO Hand empowers the operator to utilize their own tactile feedback to determine appropriate force, position, and actions, resulting in more accurate imitation of the expert's actions. We develop both non-learning and visual behavior cloning based controllers allowing HIRO Hand successfully achieves grasping and in-hand manipulation ability.

• The paper introduces the HIRO Hand, a 15-DOF wearable robotic device that uses hand-over-hand imitation learning to enhance dexterous manipulation.
• It employs both a PID controller and CNN-based behavior cloning to capture human joint movements and perform complex in-hand coordination tasks.
• Repeatability tests show a sub-0.14mm deviation under loads up to 1.5kg, achieving nearly 80% of human grasp types to ensure reliability.

Wearable Robotic Hand for Hand-over-Hand Imitation Learning

The paper presents the HIRO Hand, a wearable robotic hand designed to facilitate hand-over-hand imitation learning for dexterous manipulation. This approach seeks to address the limitations of existing data collection methods such as data gloves and virtual reality, which often lack precise tactile feedback and present difficulties in mapping human hand motions to robotic hands.

Design and Features of the HIRO Hand

The HIRO Hand is a 15-degree-of-freedom (DOF) dexterous robotic hand that is both affordable and functionally capable, offering a cost-effective solution at approximately USD 400. It is fully 3D-printed, utilizing tendon-driven mechanics to closely mimic human hand structures, thus allowing for more anthropomorphic movements. The mechanical system comprises a palm and five fingers, where each finger replicates human anatomy with multiple joints and linkages. This configuration facilitates effective grasping and manipulation tasks, making it apt for a broad spectrum of applications.

Repeatability and Test Results

The repeatability tests, conducted with varying payloads, emphasize the hand's robustness. It demonstrated a standard deviation lower than 0.14 mm for loads up to 1.5 kg across multiple cycles, indicating its reliability in executing repeated tasks with minimal deviation. The HIRO Hand can also accomplish a substantial portion of reported human grasp types (~80%).

Imitation Learning and Control Strategies

Two control strategies are implemented: a PID-based controller and a behavior cloning framework based on visual inputs:

1. PID Controller: This method utilizes a proportional-integral-derivative controller to modulate the hand's motions precisely. The hand-over-hand imitation technique records desired joint positions as a human operator manually guides the robotic hand, capturing real-time joint angle movements and employing the PID controller to replicate the actions.
2. Visual Imitation Learning: Leveraging behavior cloning, the system uses convolutional neural networks (CNNs) to learn object manipulation directly from visual inputs. Training involves input from video demos and corresponding joint configurations to predict motor actions for task replication. This approach enabled successful completion of varied manipulation tasks, including object grasping and complex in-hand coordination exercises like unscrewing a faucet.

Practical Implications and Future Directions

The HIRO Hand offers significant advancements in wearable robotic hand systems, combining cost efficiency and manipulation dexterity, which could catalyze new paradigms in robotic assistive devices and industrial automation. Its ability to integrate seamlessly with human demonstration enhances its potential for real-world applications, extending to scenarios where tactile feedback and nuanced control are paramount.

Future developments could involve enhancing the system's perceptual capabilities through advanced sensor integration, enabling finer tactile feedback and more adaptive interactions. Additionally, deploying the HIRO Hand on robotic arms for more complex task execution marks a logical progression, potentially broadening its utility in diverse sectors from healthcare to service robotics.

Conclusion

The introduction of the HIRO Hand represents a substantial step forward in the domain of dexterous robotic systems, bridging the gap between human-like dexterity and robotic efficiency. By addressing key limitations of predecessor methods and offering versatile control strategies, this system elevates the standard for robotic manipulation and presents expansive opportunities for both research and practical deployment.