In the test phase the RNN was used to simulate both online and offline conditions. The task was repeated with the iCub simulator a total of 30 times, 10 for each direction, starting from random initial positions of the joints. The action is concluded when the RNN issues the command to release the object and then errors are calculated as the difference between actual joints' positions and the reference positions. Table I summarizes the results with average, minimum and maximum error percentages. Results show that the RNN is able to learn how to control the execution of acceleration and release phases of the task. Three series of examples (one for each direction) are used as training set. Indeed the RNN coordinates with excellent precision joints' movements in order to reach their reference positions at same time, compensating the random start positions. Table II summarizes results obtained with ballistic control, showing the good ability of the RNN to control the robot also in offline mode, when orders are given according only to predictions made using the internal model. Predicted errors were calculated using output values of RNN, while Actual errors are calculated using real joints' positions of iCub. In a deeper analysis we noted that Actual error is mainly due to latencies of motors during acceleration/deceleration, that are because the prediction error in intermediate time-steps. In fact intermediate error is similar to the final one but with higher variance. Despite the good result, an higher actual than predicted error in joints' positions suggests that an additional module is needed in our architecture to incorporate lower level motor behaviours and, thus, to close the gap between the positions predicted by the network and actual ones implemented by motors. Finally we remark that these results confirm the findings of our previous work about the generalization and prediction capability in offline condition. © 2012 IEEE.

Recurrent neural network for ballistic actions: A study with the iCub

Di Nuovo A G
Primo
;
2012-01-01

Abstract

In the test phase the RNN was used to simulate both online and offline conditions. The task was repeated with the iCub simulator a total of 30 times, 10 for each direction, starting from random initial positions of the joints. The action is concluded when the RNN issues the command to release the object and then errors are calculated as the difference between actual joints' positions and the reference positions. Table I summarizes the results with average, minimum and maximum error percentages. Results show that the RNN is able to learn how to control the execution of acceleration and release phases of the task. Three series of examples (one for each direction) are used as training set. Indeed the RNN coordinates with excellent precision joints' movements in order to reach their reference positions at same time, compensating the random start positions. Table II summarizes results obtained with ballistic control, showing the good ability of the RNN to control the robot also in offline mode, when orders are given according only to predictions made using the internal model. Predicted errors were calculated using output values of RNN, while Actual errors are calculated using real joints' positions of iCub. In a deeper analysis we noted that Actual error is mainly due to latencies of motors during acceleration/deceleration, that are because the prediction error in intermediate time-steps. In fact intermediate error is similar to the final one but with higher variance. Despite the good result, an higher actual than predicted error in joints' positions suggests that an additional module is needed in our architecture to incorporate lower level motor behaviours and, thus, to close the gap between the positions predicted by the network and actual ones implemented by motors. Finally we remark that these results confirm the findings of our previous work about the generalization and prediction capability in offline condition. © 2012 IEEE.
2012
978-1-4673-4964-2
Ballistics; Recurrent neural networks; Robotics
File in questo prodotto:
File Dimensione Formato  
DiNuvo_Recurrent-Neural-Network.pdf

solo utenti autorizzati

Tipologia: Versione Editoriale (PDF)
Licenza: NON PUBBLICO - Accesso privato/ristretto
Dimensione 181.53 kB
Formato Adobe PDF
181.53 kB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.11769/738915
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
social impact