Knowledge acquisition and learning in unstructured robotic assembly environments I. Lopez-Juarez a, * , M. Howarth b a CIATEQ A.C., Centro de Tecnologia Avanzada, Manantiales 23A, Fracc. Ind. B.Q., CP 76246 El Marques, Queretaro, Mexico b Sheffield Hallam University, School of Engineering, Sheffield s1 1WB, England, UK Received 4 July 2001; received in revised form 8 October 2001; accepted 28 November 2001 Abstract Mechanical assembly by robots has traditionally depended on simple sensing systems and the robot manufacturers programming language. However, this restricts the use of robots in complex manufacturing operations. An alternative to robot programming is the creation of self-adaptive robots based on the adaptive resonance theory (ART) artificial neural network (ANN). The research presented in this paper shows how robots can operate autonomously in unstructured environments. This is achieved by providing the robot with a primitive knowledge base (PKB) of the environment. This knowledge is gradually enhanced on- line based on the contact force information acquired during operations. The robot resembles a blindfold person performing the same task since no informa- tion is provided about the localisation of the fixed assembly component. The design of a novel neural network controller (NNC) based on the Fuzzy ARTMAP network and its implementation results on an industrial robot are presented, which validate the ap- proach. Ó 2002 Elsevier Science Inc. All rights reserved. Keywords: Robotic assembly; Neural network controller; Force control; On-line learning; Adaptive resonance theory; Skill acquisition; Knowledge discovery Information Sciences 145 (2002) 89–111 www.elsevier.com/locate/ins * Corresponding author. Fax: +52-442-216-9963. E-mail addresses: ilopez@ciateq.mx (I. Lopez-Juarez), m.howarth@shu.ac.uk (M. Howarth). 0020-0255/02/$ - see front matter Ó 2002 Elsevier Science Inc. All rights reserved. PII:S0020-0255(02)00225-6