Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust Collision Detection for Robots with Variable Stiffness Actuation by Using MAD-CNN: Modularized-Attention-Dilated Convolutional Neural Network

Zhenwei Niu, Lyes Saad Saoud, Irfan Hussain

TL;DR

MAD-CNN tackles collision detection for robots with variable stiffness actuators (VSAs) by integrating a dual inductive bias—modularized joint networks and dilated convolution—with a self-attention module to achieve data-efficient, robust performance. Trained on only four minutes of collision data at the highest stiffness, the model delivers 100% detection across stiffness levels with a mean delay around $12.0503$ ms, and benefits further from a 15 ms continuous filter to reduce false positives. Ablation results show each component (modularization, dilation, attention) contributes to improved accuracy and speed, while time-windowed inputs and per-joint processing enhance data efficiency in imbalanced collision data regimes. The approach outperforms state-of-the-art methods under limited training data, offering a practical, real-time safety solution for pHRI with VSAs and enabling safer human-robot collaboration in industrial and daily tasks.

Abstract

Ensuring safety is paramount in the field of collaborative robotics to mitigate the risks of human injury and environmental damage. Apart from collision avoidance, it is crucial for robots to rapidly detect and respond to unexpected collisions. While several learning-based collision detection methods have been introduced as alternatives to purely model-based detection techniques, there is currently a lack of such methods designed for collaborative robots equipped with variable stiffness actuators. Moreover, there is potential for further enhancing the network's robustness and improving the efficiency of data training. In this paper, we propose a new network, the Modularized Attention-Dilated Convolutional Neural Network (MAD-CNN), for collision detection in robots equipped with variable stiffness actuators. Our model incorporates a dual inductive bias mechanism and an attention module to enhance data efficiency and improve robustness. In particular, MAD-CNN is trained using only a four-minute collision dataset focusing on the highest level of joint stiffness. Despite limited training data, MAD-CNN robustly detects all collisions with minimal detection delay across various stiffness conditions. Moreover, it exhibits a higher level of collision sensitivity, which is beneficial for effectively handling false positives, which is a common issue in learning-based methods. Experimental results demonstrate that the proposed MAD-CNN model outperforms existing state-of-the-art models in terms of collision sensitivity and robustness.

Robust Collision Detection for Robots with Variable Stiffness Actuation by Using MAD-CNN: Modularized-Attention-Dilated Convolutional Neural Network

TL;DR

MAD-CNN tackles collision detection for robots with variable stiffness actuators (VSAs) by integrating a dual inductive bias—modularized joint networks and dilated convolution—with a self-attention module to achieve data-efficient, robust performance. Trained on only four minutes of collision data at the highest stiffness, the model delivers 100% detection across stiffness levels with a mean delay around ms, and benefits further from a 15 ms continuous filter to reduce false positives. Ablation results show each component (modularization, dilation, attention) contributes to improved accuracy and speed, while time-windowed inputs and per-joint processing enhance data efficiency in imbalanced collision data regimes. The approach outperforms state-of-the-art methods under limited training data, offering a practical, real-time safety solution for pHRI with VSAs and enabling safer human-robot collaboration in industrial and daily tasks.

Abstract

Ensuring safety is paramount in the field of collaborative robotics to mitigate the risks of human injury and environmental damage. Apart from collision avoidance, it is crucial for robots to rapidly detect and respond to unexpected collisions. While several learning-based collision detection methods have been introduced as alternatives to purely model-based detection techniques, there is currently a lack of such methods designed for collaborative robots equipped with variable stiffness actuators. Moreover, there is potential for further enhancing the network's robustness and improving the efficiency of data training. In this paper, we propose a new network, the Modularized Attention-Dilated Convolutional Neural Network (MAD-CNN), for collision detection in robots equipped with variable stiffness actuators. Our model incorporates a dual inductive bias mechanism and an attention module to enhance data efficiency and improve robustness. In particular, MAD-CNN is trained using only a four-minute collision dataset focusing on the highest level of joint stiffness. Despite limited training data, MAD-CNN robustly detects all collisions with minimal detection delay across various stiffness conditions. Moreover, it exhibits a higher level of collision sensitivity, which is beneficial for effectively handling false positives, which is a common issue in learning-based methods. Experimental results demonstrate that the proposed MAD-CNN model outperforms existing state-of-the-art models in terms of collision sensitivity and robustness.
Paper Structure (16 sections, 9 equations, 8 figures, 2 tables)

This paper contains 16 sections, 9 equations, 8 figures, 2 tables.

Table of Contents

  1. Introduction
  2. The Proposed MAD-CNN Model
  3. Network Structure
  4. Network Structure Details
  5. Data Collection and Network Training
  6. Results
  7. Ablation Study
  8. Modularization
  9. Dilated Convolution
  10. Attention Module
  11. False Positive Handling with Continuous Filter
  12. Performance Comparison with Existing Methods
  13. CONCLUSIONS
  14. Acknowledgements
  15. Author contributions
  16. ...and 1 more sections

Figures (8)

  • Figure 1: The proposed MAD-CNN architecture for robot collision detection encompasses a comprehensive structure aimed at improving collision sensitivity, robustness, and data efficiency for robots equipped with variable stiffness actuation. MAD-CNN integrates a dual inductive bias mechanism that incorporates joint modularization and a dilated convolutional neural network, along with an attention module. It is well suited for scenarios involving hard-to-collect and imbalanced collision data.
  • Figure 2: Experiments platform. (a) A robotic manipulator with discrete variable stiffness actuators (DVSAs). (b) The working principle of the DVSA. It has four levels of intrinsic physical stiffness, and the stiffness can be rapidly adjusted online by controlling the engagement of the springs. (c) Joint Stiffness Levels of DVSA
  • Figure 3: Three cases are considered regarding joint space constraints: (a) right-side constraints to prevent contact with the table, (b) left-side constraints for avoiding table contact, and (c) constraints aimed at preventing self-collision. (d) Random point-to-point acceleration-deceleration movements without table contacts and self-collision. (e) The designed collision tool, which mainly incorporates the limit switch, is used to collide with the manipulator and record the ground truth collision labels. It outputs binary results: 1 - collision, 0 - no collision.
  • Figure 4: The time-series signals $\bold{x}(t)$ are segmented by dividing them into consecutive non-overlapping windows of size 101, with a sampling interval of 10. These segmented windows of $\bold{x}(t)$ are then sampled and arranged as the input data fed into the network.
  • Figure 5: Explanation of Detection Delay (DD), Detection Failure (DF) and False Positive (FP). The dots represent the samples
  • ...and 3 more figures