Predicting fruit-pick success using a grasp classifier trained on a physical proxy
Predicting fruit-pick success using a grasp classifier trained on a physical proxy
Embedded Files
We collected data on a system designed to emulate apple trees and trained apple pick classifiers to and tested them on apple picks on the real world.
We collected data on a system designed to emulate apple trees and trained apple pick classifiers to and tested them on apple picks on the real world.
We were able to train a pick success classifier with an AUC of 0.9 on real pick data based entirely on data collected on our apple tree proxy. This proxy allows us to collect far pick data far more quickly and with more control than field trials would normally allow.
Our pick success classifier was an LSTM, and we were also able to use it to evaluate which in hand sensors were useful for pick classification. This was presented at IROS 2022. https://ieeexplore.ieee.org/document/9981716
I trained the recurrent neural networks we used to classify a pick as either successful or failed based on simple data from in-hand sensors.
I trained the recurrent neural networks we used to classify a pick as either successful or failed based on simple data from in-hand sensors.
Page updated
Google Sites
Report abuse