Towards a Dynamic Shapley Value-Based Evaluations for Autonomous Robotic Learning from Videos

Imitation learning presents a promising pathway towards the ambitious goal of achieving autonomous task learning in robotics directly from videos. Despite notable advancements in robotic imitation learning performance, human intervention remains a crucial component in task input and performance evaluation. To realise fully autonomous imitation learning, we introduce an innovative framework that conceptualises robotic imitation learning as a dynamic cooperative game process. By integrating a module based on optical flow analysis, the framework is designed to autonomously segment complex tasks based on the stages of action dynamics. Upon completion of each stage, evaluation feedback is obtained from both the \textit{actions} and the \textit{overarching task objectives}. In addition, we incorporate the Shapley value to dynamically modulate the evaluation weights for these two dimensions contingent on the action stages. This proposed framework aims to not only accomplish full automation but also potentially bolster learning performance for complex multistage tasks. Importantly, the autonomous assignment of evaluation weights predicated on game theory is designed to allow the learning process to self-adjust the action evaluation system in response to varying task configurations and environmental changes. This adaptability enables the framework to adapt aptly to diverse working conditions, thus potentially enhancing its generalisability and transferability.