Vol.:(0123456789) 1 3 International Journal of Intelligent Robotics and Applications https://doi.org/10.1007/s41315-022-00236-0 REGULAR PAPER Development of improved coyote optimization with deep neural network for intelligent skill knowledge transfer for human to robot interaction Mahendra Bhatu Gawali 1  · Swapnali Sunil Gawali 1 Received: 21 December 2021 / Accepted: 28 April 2022 © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2022 Abstract New control approaches are being developed to allow robots to undertake increasingly dynamic and dextrous control tasks. Since these abilities need a large amount of investigation for reinforcement learning (RL), they are frequently acquired by imitation learning from physical demonstration. The cost related to the manual demonstration and its inability to scale has prompted research towards skill generalization, via contextual policies and alternatives. Despite promising outcomes, current research in this area is confned to generalizing across variations of a single ability, like throwing an object to distinct places. Modeling a robot system capable of thinking and learning has progressively been a research priority in the robotics profes- sion. Skill Transfer Learning, or the capacity to transfer human abilities to robots, has recently been a hot topic in autono- mous robotics and human–robot collaboration research. The main intention of this paper is to design and implement a novel “Transfer Expert Reinforcement Learning (TERL)” for efective skill knowledge transfer within humans and computers. Here, the modifed RL is adopted for the robotic arm movement. The involvement of improved Coyote Optimization Algorithm (COA) called Best, and Worst Fitness-based COA (BWF-COA) is used for tuning the action features of RL. As another contribution, the considered movement of the robot is determined by the deep learning model termed deep neural network with the input kinematic movements. The major aim of the modifed RL with BWF-COA is to maximize the reward, thus reducing the error diference within the desired and the predicted movement. When compared to traditional models, the results indicate that the introduced systems can beneft from signifcant information. Keywords Intelligent skill knowledge transfer · Human to robot interaction · Deep neural network · Best and Worst Fitness- based Coyote Optimization Algorithm · Reward maximization · Error minimization 1 Introduction Robots are currently needed to learn many policies inde- pendently to do various jobs in a changing environment (Kober and Peters 2009). To assist the robot to learn abili- ties from its self experiences, several learning methodolo- gies have been ofered. Owing to the complicated dynamic system, developing a meaningful control policy for robot autonomous operation is still very difcult (Mülling et al. 2013). Robots can successfully learn manipulation skills from human teaching to execute tasks. Demonstration pro- gramming has various benefts over existing programming techniques, and it can be assumed when programming by demonstration designs, which will signifcantly improve task completion success (Bennewitz et al. 2005; Kupcsik et al. 2013). Knowledge transfer is another successful model. In the areas of multitask learning and control, it is used (Kupc- sik et al. 2014). The knowledge transfer method reuses and generalizes information regarding the robot's ability. As a result, the quick adaption of the learning technology to a novel task is achieved. The approaches that were available in this context mostly focused on generalizing the user-defned trajectory into a fresh policy (Argall et al. 2009). A mapping function is created from the contextual to the policy variable, * Mahendra Bhatu Gawali gawalimahendrait@sanjivani.org.in Swapnali Sunil Gawali gawaliswapnaliit@sanjivani.org.in 1 Department of Info. Tech, SRES’s Sanjivani College of Engineering, Kopargaon, SavitribaiPhule Pune University, Pune, MS, India