Enhancing robotics learning using imitation learning through visual-based behaviour cloning

Abstract

The development of the behaviour cloning technique allows robots to mimic human experts’ behaviour by observation. The technique is mainly based on model architecture’s design and associated training mechanisms. İt is believed that such an approach will impact the importance of robotics applications in the coming future. The ongoing research presented in this paper has investigated the use of behaviour cloning with image and video data streaming to improve robot learning using imitation of human experts’ behaviour. The investigation has focused on the methodology, algorithms, and challenges associated with training robots to imitate human actions solely based on visual data inputs. An overview of the process of collecting diverse and annotated image and video datasets depicting various human actions and behaviours is presented. To provide efficient and consistent data representation, the preprocessing process includes feature extraction using convolutional neural networks (CNN) and normalization techniques. The CNN model for learning action mappings from visual inputs is described. These models’ training focuses on optimization algorithms and loss functions. A thorough examination of data quality, overfitting, and model generalization issues is addressed and presented. The research’s initial results showed the effectiveness of image and video-based behaviour cloning and how it is leading to more sophisticated and adaptive robotic systems. The limitations of the research are also discussed and presented in this paper.

Keywords

• behaviour cloning
• imitation learning
• cnn
• visual data
• robotics