Implementation of a Gaussian process-based machine learning grasp predictor

Analytic grasp success prediction with tactile feedback

“…A classification performance metric is given by the normalized area under the curve (AUC), which is 0.82 in our case. When thresholding q * to obtain 5 % false positives, we achieve a 90 % success rate (defined as the ratio of true positive rate to the sum of true positive and false positive rates [14]). …”

“…Using supervised learning and depending on varying assumptions on the availability of simulated/real data and target object information, the works in [15], [2] report overall classification accuracies between 64.6 % and 84.6 %. In [14], an AUC of 0.81 and a success rate of 88 % are achieved by using Gaussian Processes. The authors of [10] report correlation coefficients between experiments and simulations to be around 0.8.…”

“…If training data is available, it can be used to reduce the negative impact that unmodeled controller dynamics have on our approach. As a first step, we plan to build a probabilistic model for grasp success prediction [15], [14] involving the same task-oriented criterion [6] which we use in the presented work.…”

“…The underlying feature representations and machine learning methods have been evaluated by Laaksonen et al [11] with respect to their performance in determining grasp stability. Most existing approaches build probabilistic classifiers based on various grasp quality metrics [12], [13], [14]. Recent works also include features derived from tactile feedback.…”

Grasp quality evaluation done right: How assumed contact force bounds affect Wrench-based quality metrics

“…However, analyzing dynamic stability is difficult in practice as it requires to include local contact curvature, mechanical compliance, hand control laws (a compliant controller which allows hand-relative object movement results in a change of GWS) as well as the magnitude and arrangement of the applied contact forces [13], [14]. Nevertheless, our previous work [15] showed that realistic contact force modeling, involving tactile feedback and joint effort limits, allows wrench-based reasoning to predict grasp success surprisingly well and on-par with recent approaches based on supervised learning [9], [16], [17].…”