Abstract
This paper proposes a proprioceptive collision detection algorithm based on Gaussian Regression. Compared to sensor-based collision detection and other proprioceptive algorithms, the proposed approach has minimal sensing requirements, since only the currents and the joint configurations are needed. The algorithm extends the standard Gaussian Process models adopted in learning the robot inverse dynamics, using a more rich set of input locations and an ad-hoc kernel structure to model the complex and non-linear behaviors due to frictions in quasi-static configurations. Tests performed on a Universal Robots UR10 show the effectiveness of the proposed algorithm to detect when a collision has occurred.
Introduction
In human-robot collaboration (HRC), ensuring the safety of human operators is a critical requirement. Robots must be capable of detecting collisions instantaneously to mitigate potential injuries. While external solutions like tactile skins or 6-axis force/torque sensors are available, they often introduce significant costs, are prone to environmental interference, and can be fragile in industrial settings.
Proprioceptive detection—using the robot's internal sensors—offers a robust and cost-effective alternative. However, this method relies heavily on the accuracy of the robot's dynamic model. Traditional physics-based models often struggle to account for complex, non-linear friction (stiction) and internal controller behaviors, particularly in "quasi-static" scenarios where the robot moves very slowly or remains stationary. In these phases, model inaccuracies often lead to false positives in collision detection.
Methodology
The researchers developed a data-driven strategy based on Semi-Parametric Gaussian Process Regression (GPR) to overcome these modeling challenges.
Current Estimation and Residuals
Rather than calculating external torques directly, the algorithm learns to predict the expected motor current ($\hat{i}$) for any given state. Collision detection is then performed by monitoring the residual—the difference between the measured current ($i$) and the predicted current ($\hat{i}$). If this residual exceeds a predefined threshold, a collision is flagged.
Augmented Feature Vectors
Standard dynamic models typically use joint position ($q$), velocity ($\dot{q}$), and acceleration ($\ddot{q}$) as inputs. The authors identified that in quasi-static states, these kinematic features are insufficient because motor current is heavily influenced by the internal feedback controller attempting to overcome friction. To address this, they augmented the feature vector with:
- Joint position and velocity errors: Capturing proportional and derivative control actions.
- Required control currents: Capturing non-linear control actions like saturation or integral terms.
The Switching Kernel
To handle the transition between kinetic friction (moving) and static friction (nearly stopped), the model employs a specialized composite kernel. A scaling function acts as a switch:
- Kinetic Component: A standard GPR component active during general dynamics.
- Static Component: A specialized component that activates only when joint velocity falls below a threshold ($\sigma_v$), utilizing the augmented features to model unpredictable low-speed friction behavior.
Experimental Results
The approach was validated using a Universal Robots UR10. Two datasets were used: $D_1$ (random exploration for training) and $D_2$ (a circular path to test generalization). The proposed model ($SP_P$) was compared against a purely parametric model ($P_f$) and a standard semi-parametric GPR ($SP_S$).
As shown in the performance metrics, the proposed $SP_P$ model significantly reduced the Normalized Mean Squared Error (nMSE) in quasi-static configurations across almost all joints. Crucially, while standard GPR ($SP_S$) performed well on data similar to its training set, its accuracy dropped on the unseen trajectories of $D_2$. The $SP_P$ model maintained high accuracy, demonstrating its ability to learn the underlying physics more effectively.
Real-World Validation
In live tests, a human operator interacted with the UR10 arm by pushing it during both moving and nearly stationary phases. The algorithm successfully detected all four interactions with zero false positives.
Conclusion and Future Work
This research demonstrates that high-fidelity collision detection is achievable using only internal data (currents and encoders) already present in the robot's control bus. By incorporating controller-aware features and a switching kernel, the system effectively manages the complexities of quasi-static friction.
Future work will focus on automating the selection of detection thresholds and integrating the detection signal with active reaction strategies, such as admittance control or emergency stops, to provide a complete safety solution for collaborative robotics.
Summary updated