Task-Constrained Manipulation Planning in Robot Joint Space Using Long Short-Term Memory Networks

This study delves into comprehensive randomized trajectory planning within the robot joint-space framework, specifically tailored for articulated mechanisms constrained by task-space requirements. A novel representation of constrained motion is formulated to facilitate joint-space planning, leveraging Long Short-Term Memory (LSTM) networks as a cornerstone methodology. These networks adeptly encapsulate temporal dependencies and non-linear dynamics, enabling robust trajectory prediction and constraint adherence. The framework introduces two pioneering approaches for sampling joint configurations: tangent-space sampling (TS) and first-order retraction (FR), both designed to enhance global sampling efficiency for linear task-space transformations. Theoretical analysis substantiates the efficacy of FR in ensuring convergence to globally optimal solutions. This methodology addresses real-world applications encompassing workspace-coordinated tasks, such as precise rotational movements, guided linear translations, or maintaining stability under transport-induced perturbations. Furthermore, the joint-space approach effectively exploits redundant degrees of freedom (DOFs), ensuring obstacle avoidance and auxiliary goal satisfaction during task execution. Comparative evaluations reveal that the proposed methods, underpinned by LSTM networks, exhibit superior adaptability and reduced sensitivity to parametric variations relative to existing paradigms.