Sensorless Estimation of Contact Using Deep-Learning for Human-Robot Interaction

TL;DR

This work tackles sensorless external torque estimation for human-robot interaction by addressing static friction hysteresis through a Maxwell Slip-inspired Motion Discriminator and a Hierarchical Residual Learning network. The approach combines long-term temporal awareness with fast, parallel hierarchies to achieve accurate torque residuals at the current level, suitable for admittance and wrench-based control. Key contributions include the MD input scheme, the HRDL-MD architecture, and comprehensive data collection encompassing hysteresis-rich dynamics, validated in joint and task compliance settings. Practically, the method reduces reliance on expensive torque sensors while improving safety and responsiveness in collaborative robotics, with potential for adaptation to various end-effectors via calibration.

Abstract

Physical human-robot interaction has been an area of interest for decades. Collaborative tasks, such as joint compliance, demand high-quality joint torque sensing. While external torque sensors are reliable, they come with the drawbacks of being expensive and vulnerable to impacts. To address these issues, studies have been conducted to estimate external torques using only internal signals, such as joint states and current measurements. However, insufficient attention has been given to friction hysteresis approximation, which is crucial for tasks involving extensive dynamic to static state transitions. In this paper, we propose a deep-learning-based method that leverages a novel long-term memory scheme to achieve dynamics identification, accurately approximating the static hysteresis. We also introduce modifications to the well-known Residual Learning architecture, retaining high accuracy while reducing inference time. The robustness of the proposed method is illustrated through a joint compliance and task compliance experiment.

Figures (9)

• Figure 1: Snapshot of joint compliance (left) and task compliance (right). A video clip is available in the supplementary or at: https://youtu.be/Yrjf5tU94e8 for higher resolution.
• Figure 2: (a) The MLP example featuring two hidden layers, the notations $x$, $y$, $n$, and $M$ represent the input joint states, estimated currents, index of the instantaneous time step, and predefined short-term memory length, respectively. (b) The LSTM example with two LSTM cells, the symbol $h$ denotes the hidden states.
• Figure 3: The block diagram for the Motion Discriminator input scheme. TH($\cdot$) and $t_i$ denote the thresholding functions and velocity thresholds.
• Figure 4: The Maxwell Slip model of $I$ elasto-slide elements.
• Figure 5: (a) The classical Residual Learning (RDL) structure. $\tilde{x}$ represents the combined input. (b) The Hierarchical Residual Learning (HRDL) architecture. The dashed and solid lines indicate the data flow in the training phase and during online inference, respectively. The superscripts of MD, e.g., ($I_1,I_2$), indicate the subset of MD with index from $I_1$ to $I_2$.