Learning In-Hand Translation Using Tactile Skin With Shear and Normal Force Sensing (2407.07885v1)

Abstract: Recent progress in reinforcement learning (RL) and tactile sensing has significantly advanced dexterous manipulation. However, these methods often utilize simplified tactile signals due to the gap between tactile simulation and the real world. We introduce a sensor model for tactile skin that enables zero-shot sim-to-real transfer of ternary shear and binary normal forces. Using this model, we develop an RL policy that leverages sliding contact for dexterous in-hand translation. We conduct extensive real-world experiments to assess how tactile sensing facilitates policy adaptation to various unseen object properties and robot hand orientations. We demonstrate that our 3-axis tactile policies consistently outperform baselines that use only shear forces, only normal forces, or only proprioception. Website: https://jessicayin.github.io/tactile-skin-rl/

Learning In-Hand Translation Using Tactile Skin With Shear and Normal Force Sensing

In the field of dexterous manipulation, the integration of tactile feedback, especially through simulated sensors for reinforcement learning (RL), presents new avenues for improved robotic control policies. The paper "Learning In-Hand Translation Using Tactile Skin With Shear and Normal Force Sensing" introduces a novel tactile sensor model for sim-to-real transfer and details its application in developing RL policies for in-hand translation tasks using three-axis tactile skin.

Overview of Contributions

The authors have proposed a tactile skin model capable of zero-shot sim-to-real transfer while outputting ternary shear forces and binary normal forces. By leveraging this model, they developed an RL policy that receives both tactile and proprioceptive feedback. This policy is particularly designed for dexterous in-hand translation, a task requiring nuanced control over sliding contacts. The paper establishes three pivotal contributions:

1. Development of an RL-tractable sensor model for compliant tactile skin, enabling zero-shot sim-to-real transfer of tactile signals.
2. Utilization of this sensor model to train RL policies for performing in-hand translation, a highly contact-rich dexterous task.
3. Validation through extensive real-world experiments, demonstrating the superior performance and adaptability of the proposed approach compared to various baseline models.

Key Results

The performance of the proposed sensor model and its corresponding control policies was evaluated through a series of in-domain (ID) and out-of-domain (OOD) experiments. A standout observation was the consistent superior performance of policies utilizing three-axis tactile feedback over those relying purely on proprioception or simplified touch sensing.

In-Domain and Out-of-Domain Evaluation

In-Domain Cylinders: Initial experiments with uniform cylindrical objects showed that the three-axis tactile policy achieved a 38% increase in translation distance and a 94% increase in object velocity compared to proprioception-only policies. This underscores the added value of tactile feedback in controlled environments.

Out-of-Domain Tests with Tilted Hand: When the hand was tilted at various angles, tactile policies continued to outperform proprioception-only policies. The three-axis feedback maintained effective translations even as gravitational challenges due to tilting increased.

Generalization to Various Objects: Experiments with objects featuring varying geometries and dynamics, such as a hammer with a skewed center of mass (COM) and a water bottle with a variable COM, demonstrated the robustness of the tactile RL policy. Particularly, three-axis tactile policies generally achieved higher success rates and better adaptation to unseen object properties and hand tilts. For example, the S3-Axis policy consistently outperformed other models, indicating robust handling of diverse real-world conditions.

Theoretical and Practical Implications

The practical implications of this research are significant for robotic applications where adaptability and precision are paramount. Tasks involving delicate manipulation or those requiring fine adjustments in grip and orientation can greatly benefit from the enhanced control facilitated by the proposed tactile feedback system.

From a theoretical perspective, this work extends the understanding of tactile sensing in robotics, highlighting the critical role of shear forces in dexterous manipulation tasks. It also underscores the importance of high-fidelity simulations in bridging the gap between virtual training environments and real-world deployment.

Future Trajectories

Several potential avenues for future work are apparent:

• Integration with Vision: While this work focuses on tactile and proprioceptive feedback, integrating visual input could further enhance the accuracy of in-hand manipulation tasks, especially for tasks requiring precise goal specification.
• Continuous Tactile Signals: Exploring methods to simulate continuous tactile signals might further improve RL policy performance by providing richer and more nuanced sensory data.
• Real-World Adaptation: Investigating domain adaptation techniques and fine-tuning the control policies with real-world tactile feedback could yield even more robust and adaptable manipulation strategies.

Conclusion

The paper presents a compelling case for the integration of three-axis tactile sensing in robotic manipulation tasks. Through thorough experimentation and robust model development, the authors demonstrate that a nuanced approach to tactile feedback can substantially improve the performance and versatility of RL-driven dexterous manipulation policies. Overall, this research lays a strong foundation for future advancements in tactile sensing and robotic manipulation.