Fine Robotic Manipulation without Force/Torque Sensor

TL;DR

This work tackles the problem of sensing and controlling end-effector forces without a dedicated 6-axis force/torque sensor by learning a neural network mapping from internal signals (joint currents and states) to the external wrench $\tau_{ext}$ or end-effector wrench $F$. The authors propose a data-centric approach, using a base FSDS for unbiased learning and task-specific fine-tuning datasets (CSDS, HGDS) to handle contact-rich scenarios, supported by an analysis of MLP, LSTM, and CNN architectures and a robust selection of an MLP with two hidden layers of 1024 neurons. They demonstrate sensorless, real-time wrench estimation across four industrial tasks, achieving higher accuracy and stability than traditional GMO-based methods, and show practical effectiveness in free-motion wrench estimation, contact-jet control, hand-guiding, and a tight-pinning insertion with 0.1 mm clearance. The results indicate that a well-curated training strategy can enable widespread deployment of force sensing and force control capabilities on existing industrial robots without extra hardware, with implications for cost, robustness, and retrofitting older robotics platforms.

Abstract

Force Sensing and Force Control are essential to many industrial applications. Typically, a 6-axis Force/Torque (F/T) sensor is mounted between the robot's wrist and the end-effector in order to measure the forces and torques exerted by the environment onto the robot (the external wrench). Although a typical 6-axis F/T sensor can provide highly accurate measurements, it is expensive and vulnerable to drift and external impacts. Existing methods aiming at estimating the external wrench using only the robot's internal signals are limited in scope: for example, wrench estimation accuracy was mostly validated in free-space motions and simple contacts as opposed to tasks like assembly that require high-precision force control. Here we present a Neural Network based method and argue that by devoting particular attention to the training data structure, it is possible to accurately estimate the external wrench in a wide range of scenarios based solely on internal signals. As an illustration, we demonstrate a pin insertion experiment with 100-micron clearance and a hand-guiding experiment, both performed without external F/T sensors or joint torque sensors. Our result opens the possibility of equipping the existing 2.7 million industrial robots with Force Sensing and Force Control capabilities without any additional hardware.

Figures (10)

• Figure 1: Snapshot of the sensorless tight pin insertion and hand-guiding experiment setup. Video demonstrations of the experiments in Section VI is available in the supplementary materials or at: https://youtu.be/spztx3GzPzc
• Figure 2: The model structure of MLP (top left), LSTM (top right) and CNN (bottom) implemented for wrench estimation
• Figure 3: Visualization of the data region for three workspaces with three different IK Classes. The transparent blue box and opaque red plane indicate the bounding boxes and location of the fixed plates, respectively. The postures of Denso show the center of the nearest joint position clusters, or IK solutions, based on which the data was collected.
• Figure 4: Velocity distribution of Joint 1, 2, and 3 in (a) FSDS; (b) CSDS shown in histograms.
• Figure 5: Comparison between the NN-estimated (red) GMO-estimated (green) and measured wrench (black) for random pre-planned free-motion trajectories.