Introduction
As robotic systems transition from isolated industrial cages to shared environments in hospitals, homes, and factories, safety becomes the paramount design constraint. Central to safe physical human-robot interaction (pHRI) is the ability of a robot manipulator to detect collisions rapidly and reliably. This monograph presents a comprehensive survey and a unified framework for collision detection, bridging the gap between classical physics-based observers and modern data-driven learning techniques.
The Collision Detection Challenge
Historically, collision detection relied on model-based observers. These systems compare the robot's expected behavior—derived from a mathematical model of its dynamics—with real-time sensor data. However, real-world dynamics are often too complex for simple models to capture perfectly. Factors such as non-linear joint friction, cable dynamics, and varying payloads introduce errors.
Conversely, purely data-driven (model-free) methods can learn these complexities but often require massive amounts of training data and lack physical interpretability. This work addresses these shortcomings by categorizing existing methods and proposing hybrid approaches that leverage the strengths of both paradigms.
Methodological Framework
The authors categorize collision detection into three primary pillars:
- Traditional Model-Based Methods: These utilize energy, velocity, or momentum observers to estimate external torques. The momentum observer is particularly noted for its robustness, as it detects collisions without requiring noisy joint acceleration data.
- Supervised Learning: Techniques such as Support Vector Machine Regression (SVMR) and 1-D Convolutional Neural Networks (CNN) are employed to map robot states directly to collision indices.
- Unsupervised Anomaly Detection: This practical approach uses Autoencoders or One-Class SVMs (OC-SVM) to learn the "normal" state of collision-free motion. Any significant deviation from this learned norm is flagged as a collision.
The Model-Compensating Detection (MCD) Framework
The core innovation presented is the Model-Compensating Detection (MCD) framework. Rather than learning robot dynamics from scratch, MCD uses a learning model to "mop up" the residual errors (e.g., friction) left by a simplified physics model. This hybrid approach significantly increases sensitivity while reducing false alarms.
Experimental Validation and Datasets
The research was validated using data from a 6-DOF open-chain robot manipulator. The experimental design focused on several key variables:
- Collision Types: The study distinguishes between "hard" collisions (impacts with rigid objects) and "soft" collisions (contact with human flesh or soft materials).
- Payload Variation: To evaluate generalization, datasets included different payloads (e.g., 0 kg vs. 3.3 kg) to test if a detector trained on one configuration could function accurately on another.
- Ground Truth Labeling: A custom tool using Force-Sensitive Resistors (FSR) was used to establish precise ground truth for training and evaluation.
Key Insights and Results
A significant finding is the efficacy of Unsupervised Learning. Autoencoders (MCD-AE) achieved performance comparable to supervised methods without requiring thousands of intentional collisions for training. By learning the "envelope" of safe operation, these models proved robust to small physical changes in the robot, such as wear and tear.
The study utilized the Area Under the Precision–Recall Curve (PRAUC) to scientifically evaluate the trade-off between detection delay and false alarm rates. Results showed that MCD-CNN and MCD-SVMR excelled in supervised settings by effectively filtering dynamic noise. Furthermore, the proposed methods successfully distinguished between the high-frequency signals of hard collisions and the lower-frequency signals of soft "shoves," which is critical for nuanced safety responses.
Limitations and Future Work
Despite the robustness of the MCD framework, challenges remain:
- Large Dynamics Changes: While the models handle minor payload variations, significant changes in configuration or environment may still necessitate retraining or the application of Transfer Learning.
- Contact-Rich Tasks: Future research will focus on complex tasks like assembly, where the system must distinguish between "intended contact" and "accidental collision."
- Multi-Robot Systems: Extending these successes to coordinated multi-arm systems represents a significant future research frontier.
Summary updated