Robotic Imitation Learning from Videos

: Boosting Autonomy and Trasnferability

Abstract

Robotic Imitation Learning, utilizing visual inputs from videos, plays a crucial role in the advancement of automated task learning for robots. Despite its clear potential, significant challenges impede its widespread application and optimal performance:
1. Feedback Integrity in Learning Phases: Current systems mostly rely on binary feedback mechanisms, which lack the detail needed to accurately improve a robot’s performance in complex tasks. This binary feedback limits the comprehensive enhancement of robotic capabilities as it hinders precise identification and improvement of specific actions within a task sequence.
2. Semantic Consistency in Task Segmentation: Determining the optimal granularity for task segmentation is crucial, as it directly impacts the efficiency of the learning phase, as well as the robustness and precision of robotic execution in dynamic environments.
3. Generalizability and Transferability: The limited adaptability of existing feedback modules restricts a robot’s versatility, especially when transitioning between different tasks in varying environments. This hampers the robot’s ability to transfer learned tasks effectively across different operational contexts, thus affecting task execution and adaptation in environments with diverse characteristics and requirements.

In addressing the identified challenges, this research introduces a new framework based on cooperative game theory, aimed at enhancing robotic imitation learning. The main contributions of this framework are as follows:
1. Feedback mechanisms are refined by incorporating insights from both immediate actions and overall task objectives. The strategic use of the Shapley Value ensures a dynamic adjustment of evaluation weights at different stages of action.
2. The framework integrates an optical flow analysis module to facilitate autonomous task segmentation. This module effectively identifies variations in actions and task objectives, enabling a dynamic and fine-grained decomposition of complex actions.
3. The framework’s foundation in game theory provides inherent adaptability. It allows the robot to autonomously adjust evaluation parameters, resulting in an architecture with improved generalizability and transferability.
Empirical evaluations confirm that the framework significantly enhances performance, especially in complex, multi-stage tasks. In summary, this thesis marks a significant advancement in fully unlocking the potential of robotic imitation learning, supported by comprehensive experimental validation.

Date of Award	2024
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Fei Chao (Supervisor) & Changjing Shang (Supervisor)