SpikeATac: A Multimodal Tactile Finger with Taxelized Dynamic Sensing for Dexterous Manipulation

TL;DR

SpikeATac presents a multimodal tactile finger that fuses a high-frequency, taxelized PVDF dynamic sensor with static capacitive sensing to enable rapid yet delicate manipulation, including in-hand handling of fragile objects. The authors fabricate and integrate the sensor into a four-finger hand and develop a learning pipeline that combines imitation learning with on-robot reinforcement learning guided by tactile rewards and human feedback. They demonstrate fast, gentle grasping on delicate objects and successful in-hand rotation of fragile items, enabled by raw sensor signals rather than sim-to-real proxies. This work highlights the practical potential of dense, multimodal tactile sensing for real-world dexterous manipulation and learning-based control.

Abstract

In this work, we introduce SpikeATac, a multimodal tactile finger combining a taxelized and highly sensitive dynamic response (PVDF) with a static transduction method (capacitive) for multimodal touch sensing. Named for its `spiky' response, SpikeATac's 16-taxel PVDF film sampled at 4 kHz provides fast, sensitive dynamic signals to the very onset and breaking of contact. We characterize the sensitivity of the different modalities, and show that SpikeATac provides the ability to stop quickly and delicately when grasping fragile, deformable objects. Beyond parallel grasping, we show that SpikeATac can be used in a learning-based framework to achieve new capabilities on a dexterous multifingered robot hand. We use a learning recipe that combines reinforcement learning from human feedback with tactile-based rewards to fine-tune the behavior of a policy to modulate force. Our hardware platform and learning pipeline together enable a difficult dexterous and contact-rich task that has not previously been achieved: in-hand manipulation of fragile objects. Videos are available at \href{https://roamlab.github.io/spikeatac/}{roamlab.github.io/spikeatac}.

Figures (8)

• Figure 1: Overview of SpikeATac. Left: completed finger and design. Right: integration on a parallel gripper and multifingered robot hand. Combining high-performance, taxelized PVDF for dynamic sensing with capacitive pads for static sensing, SpikeATac enables fast yet delicate manipulation. Used in conjunction with imitation learning and on-robot reinforcement learning fine-tuning, SpikeATac data can also enable multifingered dexterity such as in-hand reorientation even on fragile objects.
• Figure 2: The PVDF fabrication process along with a photo of the finished sensor attached to the PCB stack with an FPC cable.
• Figure 3: PVDF, capacitive, and ATI Gamma responses when probing with a hemispherical indenter ($d=6mm$, see Fig. \ref{['fig:heatmap']}) at two linear probe speeds: 1mm/s (left) and 10mm/s (right). We test when under a moderate touch condition ($\sim$3.5N), light touch condition (minimal contact established visually), and when approaching but not touching. Bottom-center PVDF and capacitive taxels are shown (location 8 on Fig. \ref{['fig:heatmap']}). ADC counts are normalized to the full measurement range; CDC counts are normalized to just above the moderate touch's maximum response. We note that, at the higher speed, PVDF can detect light contact that the other sensors cannot, or detect contact earlier than the other sensors.
• Figure 4: Heatmaps of the maximum absolute value of the zeroed PVDF response from each taxel (i.e., a dark taxel represents an increasing or decreasing signal) depending on indentation location, showing the spatial information carried by the PVDF taxels.
• Figure 5: (a) Setup of the fast and delicate grasping experiment, in which we command a slow, medium, or fast velocity to the gripper and stop after contact is detected, comparing PVDF-based and capacitive-based detection (Tab. \ref{['tab:fastgrasp']}). (b) After 15 trials, the more fragile object (seaweed) is noticeably crumpled when using capacitive-based stopping, but appears nearly untouched in the PVDF condition at fast velocity. In fact, in the medium and fast conditions, the capacitive method often failed to detect contact at all. (c) Raw signals from one finger show a distinct PVDF spike at seaweed contact, while the capacitive signal does not reliably rise above its noise floor (each line shows an individual taxel from one of the two fingers).