Opportunities and Challenges of Deep Learning in Robotic Manipulator Planning: A Survey

Deep learning modules have recently been used to improve the algorithmic primitives of sampling-based and global optimization trajectory planning algorithms. Figures are adopted from reviewed papers in the survey.

Abstract

State-of-the-art generalist manipulation policies have enabled the deployment of robotic manipulators in unstructured human environments. However, these frameworks struggle in cluttered environments primarily because they utilize auxiliary modules for motion planning and control. Motion planning remains challenging due to the high dimensionality of the robot's configuration space and the presence of workspace obstacles. Neural motion planners have enhanced motion planning efficiency by offering fast inference and effectively handling the multi-modality of the motion planning problem. Despite such benefits, current neural motion planners often struggle to generalize to unseen, out-of-distribution settings. This paper reviews and analyzes the state-of-the-art neural motion planners, highlighting both their benefits and limitations. It also outlines a path toward establishing generalist neural motion planners capable of handling domain-specific challenges.

The Struture of the Survey

Survey strucutre with references to the sections of the original paper. Figures are adopted from reviewed papers in the survey.

Planning within Manipulation Autonomy Stack

A toy example of a robotic manipulator's autonomy stack. Deep learning techniques have significantly enhanced all the components within this autonomy stack. The focus of this survey paper is exploring how deep learning modules have improved the motion planning sub-stack, either by improving specific components or classical motion planners or by functioning as end-to-end planners.

Planning Preliminaries and Classical Planning Algorithms for Robotic Manipulators

Workspace vs. configuration space for a 2-DOF planar manipulator.

Algorithmic primitives (I. Sampling, II. Steering, and III. Collision Checking) of sampling-based planning algorithms.

An abstract illustration demonstrating how global trajectory optimization techniques rely on the initial trajectory to warm-start the optimization process. This highlights that even slightly different initializations can lead to distinct final trajectories.

(a) Calculating the distance to collision via geometric collision checkers is computationally expensive. A common practice is to utilize convex geometric primitives (e.g., spheres, cylinders) to enclose manipulator links and joints for efficient collision querying. (b) Spatial decomposition (SD) methods are commonly utilized for point-cloud processing and workspace partitioning for collision detection. Figure taken from reviewed papers in the original manuscript.

Deep Learning Basics and Potential For Robotic Manipulator Planning

Common deep learning modules and their applications in planning (part 1).

Common deep learning modules and their applications in planning (part 2).

Citation

@article{soleymanzadeh2026toward,
  author    = {Soleymanzadeh, Davood and Lopez-Sanchez, Ivan and Su, Hao and Li, Yunzhu and Liang, Xiao and Zheng, Minghui},
  title     = {Toward Generalist Neural Motion Planners for Robotic Manipulators: Challenges and Opportunities},
  journal   = {IEEE Transactions on Automation Science and Engineering},
  year      = {2026},
}