Tactile-Morph Skills: Energy-Based Control Meets Data-Driven Learning

TL;DR

The paper tackles tactile, contact-rich robotic manipulation in industrial settings by proposing the Tactile-Morph Skills framework, which merges unified force-impedance control with data-driven energy distribution learning. A Temporal Convolutional Network estimates the energy required for a given trajectory and force policy, enabling an energy-tank based safety mechanism that discretely injects energy to preserve stability and prevent unsafe interactions. The nominal tactile skill is derived from Lagrangian dynamics with passivity guarantees, and the framework adapts motor and force policies (morphing) according to the estimated energy landscape to achieve zero-shot transfer across surfaces with similar friction. Experimental results show accurate energy trajectory estimation (MAPE ~7.9% for power trajectories; ~2.28% for energy sums), robust zero-shot transfer to planar and inclined geometries, and enhanced safety in real-robot deployments, highlighting practical potential for industrial polishing and similar tasks. Overall, the work delivers a data-driven, energy-aware approach to safe, transferable tactile manipulation with tangible benefits for automation, safety, and skill reuse across tasks and surfaces.

Abstract

Robotic manipulation is essential for modernizing factories and automating industrial tasks like polishing, which require advanced tactile abilities. These robots must be easily set up, safely work with humans, learn tasks autonomously, and transfer skills to similar tasks. Addressing these needs, we introduce the tactile-morph skill framework, which integrates unified force-impedance control with data-driven learning. Our system adjusts robot movements and force application based on estimated energy levels for the desired trajectory and force profile, ensuring safety by stopping if energy allocated for the control runs out. Using a Temporal Convolutional Network, we estimate the energy distribution for a given motion and force profile, enabling skill transfer across different tasks and surfaces. Our approach maintains stability and performance even on unfamiliar geometries with similar friction characteristics, demonstrating improved accuracy, zero-shot transferable performance, and enhanced safety in real-world scenarios. This framework promises to enhance robotic capabilities in industrial settings, making intelligent robots more accessible and valuable.

Figures (7)

• Figure 1: Pipeline Overview: Tactile-Morph Skills Framework is composed of an energy distribution estimator and unified force-impedance control(UFIC). During the training phase, a set of tactile skills, designed based on basic motion patterns, is planned using point cloud observations of the working surface. These tactile skills are executed by the robot, and the corresponding energy distribution is recorded in real-time. The collected data is then utilized to train our energy distribution estimation model. Upon completion of the training process, the energy distribution for any novel tactile skill sampled from the working surface can be accurately estimated. By integrating the estimated energy distribution with UFIC, we can provide a formal guarantee of passivity to any contact-rich tactile skill.
• Figure 2: Energy Distribution Estimation Pipeline.
• Figure 3: Energy Heat Map: inferring randomly generated trajectories and integrating the power distribution.
• Figure 4: Estimated Power Trajectory in comparison with ground truth (in blue).
• Figure 5: Safety Performance Comparison: Our method and baseline performances are illustrated separately in the first and second rows. The green dots indicate maintained contact, the yellow dots indicate losing contact, and the red dots represent the worst-case scenario—hitting the board. In the first row, our method injects energy step-wise into the system. The energy within the robot system remains minimal and is frequently refreshed, preventing the robot from falling and hitting the working surface. In contrast, the baseline (shown in the second row) starts with a constant energy value. As a result, the controller cannot respond rapidly to the accident falling (losing contact), owing to the remaining energy in the energy tank.