AnySkin: Plug-and-play Skin Sensing for Robotic Touch (2409.08276v3)

Abstract: While tactile sensing is widely accepted as an important and useful sensing modality, its use pales in comparison to other sensory modalities like vision and proprioception. AnySkin addresses the critical challenges that impede the use of tactile sensing -- versatility, replaceability, and data reusability. Building on the simplistic design of ReSkin, and decoupling the sensing electronics from the sensing interface, AnySkin simplifies integration making it as straightforward as putting on a phone case and connecting a charger. Furthermore, AnySkin is the first uncalibrated tactile-sensor with cross-instance generalizability of learned manipulation policies. To summarize, this work makes three key contributions: first, we introduce a streamlined fabrication process and a design tool for creating an adhesive-free, durable and easily replaceable magnetic tactile sensor; second, we characterize slip detection and policy learning with the AnySkin sensor; and third, we demonstrate zero-shot generalization of models trained on one instance of AnySkin to new instances, and compare it with popular existing tactile solutions like DIGIT and ReSkin. Videos of experiments, fabrication details and design files can be found on https://any-skin.github.io/

• The paper presents a tactile sensor that streamlines fabrication to enhance durability and signal consistency in robotic applications.
• It demonstrates 92% accuracy in slip detection using an LSTM model, underscoring its efficacy in real-world manipulation tasks.
• The sensor maintains robust cross-instance policy generalization with only a minor performance drop (~15%) when transferred to new instances.

AnySkin: Plug-and-play Skin Sensing for Robotic Touch

The paper presents "AnySkin," a versatile, self-adhering tactile sensor for robotic applications. It addresses key challenges in tactile sensing, namely versatility, replaceability, and data reusability. This research builds upon the magnetic-field-based ReSkin sensor, introducing several innovations to improve fabrication, durability, and signal consistency. The authors assert AnySkin's efficacy in detecting slip and ensuring policy generalization across different sensor instances, comparing its performance with existing tactile sensors such as ReSkin and DIGIT.

Key Contributions

Streamlined Fabrication Process and Design Improvements

AnySkin employs a streamlined fabrication process that enhances the durability and consistency of tactile sensing. The authors use a pulse magnetizer to magnetize the skins post-curing, resulting in stronger magnetic fields. This approach contrasts with the curing process under the influence of a magnetic field, as used in ReSkin. Additionally, the use of finer magnetic particles (25μm) achieves a more uniform distribution, mitigating issues like particle settling, which previously compromised signal consistency.

Characterization of Slip Detection and Policy Learning

A significant experiment involved slip detection using an LSTM model trained on tactile data collected from a Jaco robot equipped with AnySkin sensors. The results demonstrated a 92% accuracy rate in detecting slip on unseen objects. This performance underscores the sensor’s capability in practical applications, where slip detection is critical for effective manipulation.

Zero-shot Generalization of Models

The paper's bold claim is AnySkin's ability to generalize learned manipulation policies from one instance of the sensor to new instances. This is substantiated through various experiments spanning tasks like plug insertion, card swiping, and USB insertion. Policies trained using AnySkin demonstrated a minor performance drop (average of 15.6%) when transferred to new sensor instances, as opposed to a significant drop observed in ReSkin and DIGIT. Notably, this finding positions AnySkin as a scalable solution for tactile sensing in robotics.

Experimental Results

Signal Strength and Consistency

Experimental comparisons between ReSkin and AnySkin revealed that the latter exhibits superior signal strength and remarkably low variability across instances. For instance, ReSkin showed a normalized standard deviation of 0.54 in $B_{xy}$, while AnySkin demonstrated a substantially lower variability of 0.12. The self-aligning design of AnySkin eliminates the need for adhesives or screws, thus enhancing both ease of replacement and signal consistency.

Ease of Replaceability

AnySkin's ease of replaceability was validated through a user paper. Participants took an average of 12 seconds to replace the sensor, significantly outperforming ReSkin (adhesive: 82 seconds, screws: 236 seconds) and DIGIT (58 seconds). Furthermore, AnySkin skins proved reusable after replacement, unlike adhesive-based ReSkin skins.

Policy Learning and Cross-instance Generalizability

The evaluation on policy learning tasks showed that AnySkin, when swapped with a new instance, maintained a high success rate (average drop of only 13%). In contrast, policies using ReSkin and DIGIT experienced substantial degradation when tested with new instances. For example, in the plug insertion task, the success rate with a swapped ReSkin dropped from 6 to 1.7, highlighting AnySkin’s superior generalizability.

Implications and Future Developments

The practical implication of AnySkin's design and capabilities extends to scalable tactile datasets and robust deployment scenarios. The sensor's consistent signal response and ease of replaceability address current limitations in tactile-based robotic learning. As future developments, integrating AnySkin into large-scale data collection frameworks or enhancing signal consistency through machine learning approaches could further bolster the sensor's utility in real-world applications.

However, AnySkin still faces challenges such as interference from magnetic and ferromagnetic objects, albeit partially addressed by prior noise reduction techniques. Exploring calibration schemes or understanding the effects of sensor material properties could refine its performance further.

Conclusion

AnySkin represents a significant advancement in tactile sensing for robotics, combining ease of use, consistency, and versatility. Through rigorous experimentation, the authors substantiate AnySkin's potential in generalizing learned manipulation policies across sensor instances, marking a milestone in the scalability and practical deployment of tactile sensors in robotics. The paper provides a foundational framework for future research in tactile sensor development and applications.