Embedded Files
Neel Jawale*, Navneet Kaur*, Amy Santoso, Xiaohai Hu, Xu Chen#
University of Washington
*: equal contribution, #: corresponding author
accepted at IROS 2024
Abstract
Abstract
Safely handling objects and avoiding slippage are fundamental challenges in robotic manipulation, yet traditional techniques often oversimplify the issue by treating slippage as a binary occurrence. Our research presents a framework that both identifies slip incidents and measures their severity. We introduce a set of features based on detailed vector field analysis of tactile deformation data captured by the GelSight Mini sensor. Two distinct machine learning models use these features: one focuses on slip detection, and the other on evaluating the slip's severity which is the slipping velocity of the object against the sensor surface. Our slip detection model achieves an average accuracy of 92%, and the slip severity estimation model exhibits a mean absolute error (MAE) of 0.6 cm/s for unseen objects. To demonstrate the synergistic approach of this framework, we employ both of these models in tactile feedback-guided vertical sliding task. Leveraging the high accuracy of slip detection, we utilize it as the foundational and corrective model and integrate the slip severity score into the feedback control loop to address slips without overcompensating.
To evaluate the effectiveness of the slip detection and slip severity models, both individually and in combination as a learned slip detection-severity framework, we conduct three experiments. The first experiment involves deploying the developed slip detection models and assessing their performance through comparative analysis. The second experiment is dedicated to implementing and evaluating the slip severity models, utilizing the same flow features extracted from tactile deformation field for slip detection. The third experiment explores a synergistic approach to slip detection and severity assessment by integrating a feedback Proportional Derivative (PD) gripper controller for vertical object sliding downstream task.
Slip Detection
Slip Detection
Our study concentrates on using Random Forest (RF) and Gradient Boosting (GB) as our slip detection models.
The STATIC + GRASP scenario details the methodology for acquiring `no-slip' data labeled 0, while the SLIP scenario illustrates the process for gathering data indicative of slip labeled 1. Both datasets are combined during the training phase for input into the slip detection model.
In assessing the classification models' effectiveness in this study, we use the following metrics to capture various aspects of the model's efficacy.
Here, True Positives (TP) are instances correctly predicted as positive, True Negatives (TN) are instances correctly predicted as negative, False Positives (FP) are instances incorrectly predicted as positive, and False Negatives (FN) are instances incorrectly predicted as negative.
Results of trained classifiers on unseen objects
Results of trained classifiers on unseen objects
Slip Severity Estimation
Slip Severity Estimation
We employ two deep learning-based models using PyTorch: a Long Short-Term Memory (LSTM) network and a Multilayer Perceptron (MLP) . These models are chosen to explore both the sequential nature of the data and the relationships among static features at discrete time points, respectively.
The gripper, fitted with a GelSight Mini sensor, is programmed to slide over a fixed object. This setup synchronously captures tactile feedback and slip velocity data—the latter serving as ground truth—to train the neural network in assessing slip severity.
To evaluate the performance of our learned slip severity estimation models, we utilized metrics that capture different aspects of predictive accuracy and error magnitude.
Results of trained Slip Severity Estimation models on unseen objects
Results of trained Slip Severity Estimation models on unseen objects
Vertical Sliding Manipulation Control Task
Vertical Sliding Manipulation Control Task
We use Random Forest as our slip detection classifier and LSTM as our slip severity estimator given their performance and generalizability to unseen, challenging objects. These models are trained on the combined feature set to extract the best results. The object selected for this demonstration is a smooth PVC pipe, which was not included in the training set of objects. The surface geometry, texture, and inertia of the pipe are unknown to the GelSight sensor.
Page updated
Report abuse