Related Works

[jsiburian]

Relevant Papers and Datasets

Dataset

Yan Gang's Paper:

• Source Code: https://github.com/yg19930918VT/Slip_detection
• Dataset: https://zenodo.org/records/4584809

Lorenzo Jamone's paper:

• Dataset: https://github.com/ARQ-CRISP/slip_detection_dataset_2021

[jsiburian]

Leveraging distributed contact force measurements for slip detection: a physics-based approach enabled by a data-driven tactile sensor

• Venue: ICRA 2022
• Paper URL: https://arxiv.org/abs/2109.11504
• Dataset: None

Summary

• Method: Physics-based, Coulomb model and Stick Ratio (SR)
• Sensor: Vision-based

Motivation

• Vast majority of works purely based on distributed measurements have focused on data-driven approaches.
• Model-based approach to bypass dependency on training data while retaining reliable prediction accuracy.
• Novelty point: Distributed contact force measuremens are leveraged to measure the slipping area

Methods

Tactile Sensor

• Tactile sensor used: https://www.mdpi.com/1424-8220/19/4/928
• Camera frames -> DNN -> discretized force distribution
• 20x20, 400 measurement points with xyz axes measurements
• Only one sensor used for computational efficiency

Method

Baseline

• Dry friction law (Coulomb friction law)
• Computed as total quantities (summations) over the entire contact region

Stick Ratio Approach

• Based on the concept of incipient slip (relative displacement occurs only in a narrow region of the contact interface, preceding gross slip)
• Consider each taxel as a discretized contact point, compare normal and shear forces at a single point
• Ratio of stick area (C) against contact area (A) is called stick ratio (SR)

Experiment

• Constant value of friction coefficient set to 0.45 (assumption from previous work)
• Grasping sequence defined a priori for each test object, initial object pose facilitated successful grasp
• External load manually applied until slip happened
• Change of slip state (STICK or SLIP) recorded manually, ground truth labels to compute performance scores
• Label data is for evaluation only (model-based), not used for training
• Tested on 5 objects, 5 displacements = 4 translational (2 horizontal 2 vertical), 1 rotational with 90 angle

Experimental Results

• Evaluation metrics: Accuracy, Precision, Recall
• Results were obtained as average of three different runs for each experiment (reduce uncertainty of ground truth label)
• Stick ratio outperforms baseline strategy, able to handle rotational slip
• Slip = stick ratio < threshold of 0.5 (arbitrarily chosen)

Limitations

• Performance gap between translational and rotational slip -> marginal drop in accuracy of contact forces during rotational motion -> geometric interferene between the slider and the edges of the squared tactile sensor when rotating object
• Possible solution: dome-shaped tactile sensor?

Comments

• Should be straight forward to implement onto uSkin
• Only uses one sensor, limited frequency
• Still relies on threshold and assumption of friction coefficient
• Limited objects and limited testing scenario (understandable because we need to easily obtain ground truth labels)
• Would be interesting to test onto Lorenzo's dataset for example (pick-and-place)

[jsiburian]

Slip Detection for Grasp Stabilization With a Multifingered Tactile Robot Hand

Venue: IEEE Transactions On Robotics (2021)

Summary

• Sensor: TacTip (Optical biomimetic)
• Slip Method: 1 model-based + 2 learning-based (SVM & LogReg)

Main Contributions

• Compare and understand simple slip detection methods based on pin velocity features
• Test multiple slip scenarios with 11 different objects
• Testing two real-world scenarios: adding weight to destabilize a grasp & perform first-time grasping

Motivation

• Debate on model-based versus model-free is far from being settled, and merits of both approaches should be investigated.

Method

Compare three binary classifiers:

• A threshold on the magnitude of mean pin velocity
• SVM applied to either linear or nonlinear combinations of pin velocities
• logistic regression (LogReg) applied to individual pin velocities

Ground truth data:

• Data labeling: ArUco optical marker on the slipping object and recorded its height with another JeVois camera
• Reasoning: labeling process autonomous and contains no labeling bias

Note:

• For efficiency, they test the effect of downsampling these data
• Vary the number of frames in the slip class and randomly downsample the data before optimizing each classifier

Experiment (Single Finger)

Uses a rail system and single tactile fingertip Start with offline performance with pre-collected data Online performance to detect slip in real-time (best classifier only)

Offline Performance

Threshold Method

• Threshold was chosen by testing 10 values (Best performing was 0.14 pixels/s)
• F1 score = 0.753
• Significantly worse than SVM and LogReg
• Suffers from not rejecting false positives (it does not distinguish slip from changes of pin position unrelated to slip)

SVM and LogReg methods

• SVM outperforms LogReg across all tests
• Reaction time: able to classify in sub-milisecond time (therefore we can select purely on score)

Online Performance

• Test the ability to detect slip in real-time (best classifier only) -> action to minimize slip
• Two reaction strategies: any frame vs. two consecutive frames (latter is to reduce false positives at the cost of reaction time)
• Result: Similar task success, but FP rate drops from 20% to 0% when using two consecutive frames

Experiment (Whole Hand Testing)

• TO-DO READ

Evaluation

• Does not use normal method of using accuracy (Reasoning: usually more examples of nonslip than slip, but they do not want to discard data that could give a better classifier, therefore normal accuracy is not a good method)
• F1 score is used (which takes into account the precision and recall)

Comments

• Paper provides a very clear and comprehensive comparison of the slip methods, great paper overall

Interesting things to adopt:

• Compare performance of using one vs. two sensors (number of sensors detecting slip)
• Number of consecutive frames detecting slip (vary the frames, can even test with the threshold method used in Fuso internship)

Interesting Reference to Read

• [ ] L. Li, F. Sun, B. Fang, Z. Huang, C. Yang, and M. Jing, “Learning to detect slip for stable grasping,” in Proc. IEEE Int. Conf. Robot. Biomimetics, 2017, pp. 430–435
• [ ] M. Li, Y. Bekiroglu, D. Kragic, and A. Billard, “Learning of grasp adaptation through experience and tactile sensing,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Sep. 2014, pp. 3339–3346