On-Line Incremental Learning for Unknown Conditions during Assembly Operations with Robots

To be effective in real operations where the environment is continuously changing, robots have to perceive the environment and to adapt accordingly. Unfortunately, there are uncertainties due to ageing of mechanisms, isturbances, backlash, etc. that limit the usage of current control algorithms. In this paper we propose an on-line incremental learning technique using Fuzzy ARTMAP and a pattern selection criterion. The technique starts by training the ANN with a primitive knowledge base. In the presence of new patterns, the criterion-based on the success of the current action-decides autonomously if the pattern should be learned, if the ANN has to recall, or if a recovery action must be performed. The incremental learning approach is based on the online update of the neural network weights and the defined criterion decides should the new pattern be learned. The peg in hole operation (PIH) is selected as the study case in order to evaluate the performance of the technique, which is described in detail as well as the basics of the peg in hole operation. Promising results obtained with real operations with an industrial robot without over training/forgetting is presented that validate the approach.