Robot Joint Collision Detection Using Observer and LSTM Torque Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current collision detection methods for robots are inaccurate due to difficulties in establishing an accurate joint friction model and obtaining joint acceleration, leading to false collision detection and reduced safety in robot operation.
Innovation Solution
A collision detection method using a generalized momentum-based disturbance observer to calculate joint external torque and a long short-term memory (LSTM) network to learn and compensate for unmodeled errors, improving the accuracy of collision detection by reducing false positives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a threshold is set for the output torque of the joint motor or the change rate to determine collision, then the detection method is simple, but the accuracy is low and only large contact forces can be detected
Solution Approach 1:
The patent replaces the mechanical threshold-based torque detection method with an observer-based estimation system. The dynamic observer continuously estimates external torque by processing motor current and joint position data through dynamic models, enabling accurate detection of small contact forces that would be missed by fixed thresholds.
Solution Approach 2:
The patent introduces an intermediate estimation layer between the motor torque output and collision determination. The dynamic observer acts as an intermediary that processes raw motor data through friction models and dynamic equations to produce refined external torque estimates, improving detection accuracy before final collision judgment.
2Measurement precision
If a dynamics model is established to predict joint torque through parameter identification, then collision detection accuracy can be improved, but it is difficult to establish an accurate joint friction model and parameter identification errors remain
Solution Approach 1:
The patent implements feedback mechanisms where the dynamic observer continuously compares estimated torque with actual motor torque measurements. The friction model parameters are refined through iterative identification processes that use observed deviations to adjust model accuracy, reducing parameter identification errors over time.
Solution Approach 2:
The patent employs parameter identification techniques that adapt friction model parameters based on observed robot operation data. By continuously adjusting friction coefficients and inertial parameters to match actual system behavior, the model accuracy improves while managing the complexity of maintaining detailed dynamics models.
3Measurement precision
If joint acceleration is obtained through velocity differentiation to calculate joint torque, then the dynamics model can be used for collision detection, but data noise is introduced and large errors occur in joint external torque calculation
Solution Approach 1:
The patent replaces the direct differentiation method for obtaining acceleration with an observer-based estimation approach. Instead of computing acceleration by differentiating velocity signals (which amplifies noise), the dynamic observer estimates acceleration indirectly through the robot's dynamic model and measured forces, significantly reducing noise propagation.
Solution Approach 2:
The patent introduces the dynamic observer as an intermediary between velocity measurements and acceleration calculation. The observer uses the robot's dynamic model to estimate acceleration from force measurements and position data, acting as a noise-filtering intermediary that produces smoother, more accurate acceleration estimates for torque calculation.
4Reliability
If a large determination threshold is used to avoid false detection of collision, then false positives are reduced, but the accuracy of collision detection is reduced
Solution Approach 1:
The patent transitions from static threshold values to dynamic thresholds that adapt based on the robot's current operating state. The determination threshold is adjusted according to the dynamic observer's confidence level, operational mode, and environmental context, allowing high sensitivity during safe operations while maintaining reliability during uncertain conditions.
Solution Approach 2:
The patent implements adaptive threshold parameters that change based on operational context. Rather than using a fixed large threshold that reduces accuracy, the system dynamically adjusts the determination threshold based on friction model confidence, operational mode, and observed system behavior, optimizing both reliability and accuracy for different operating conditions.
Data Source
AI summary
A collision detection method, a storage medium, and a robot are provided. The method includes: calculating an external torque of a first joint of the robot based on a preset generalized momentum-based disturbance observer; calculating an external torque of a second joint of the robot based on a preset long short-term memory network; calculating an external torque of a third joint of the robot based on the external torque of the first joint and the external torque of the second joint; and determining whether the robot has collided with an external environment or not based on the external torque of the third joint and a preset collision threshold. In the present disclosure, the component of the model error in the joint external torque calculated by the disturbance observer is eliminated to obtain the accurate contact torque, thereby improving the accuracy of the collision detection.