Towards Generalist Neural Motion Planners
for Robotic Manipulators

Challenges and Opportunities

Davood Soleymanzadeh1, Ivan Lopez-Sanchez2, Hao Su2, Yunzhu Li3, Xiao Liang4, Minghui Zheng1
1 Texas A&M University, Mechanical Engineering 2 New York University 3 Columbia University 4 Texas A&M University, Civil and Environmental Engineering
IEEE Transactions on Automation Science and Engineering
Paper Paper List Explore BibTeX
Survey Website

A map of how deep learning is changing manipulator motion planning.

The survey organizes neural motion planners by where learning enters the planning stack: end-to-end planning, sampling, steering, trajectory optimization, and collision checking.

Deep learning for robotic manipulator motion planning overview

Abstract

State-of-the-art generalist manipulation policies have enabled robotic manipulators to operate in unstructured human environments. However, these frameworks still struggle in cluttered settings because they often rely on auxiliary modules for low-level motion planning and control. Motion planning remains difficult due to high-dimensional configuration spaces and workspace obstacles. Neural motion planners improve planning efficiency through fast inference and by modeling the multimodality of planning problems, but they often struggle to generalize to unseen, out-of-distribution settings. This survey reviews neural motion planners for robotic manipulators, analyzes their benefits and limitations, and outlines a path toward generalist neural motion planners.

What This Survey Gives You

From planner primitives to generalist neural planners.

Read it as an atlas: choose the planning component you care about, see what neural methods have been used, and understand the remaining gap to robust deployment.

Planning Stack

Connects workspace, configuration space, sampling, steering, optimization, and collision checking.

Neural Modules

Summarizes MLPs, CNNs, transformers, GNNs, generative models, point clouds, and neural fields.

Taxonomy

Groups reviewed work by where learning improves the planning pipeline.

Roadmap

Highlights generalization, safety, data, benchmarking, and domain-specific deployment challenges.

Survey Atlas

Structure

How the survey is organized

The paper starts from motion planning preliminaries, then moves through classical planners, deep learning modules, neural planning methods, challenges, and application domains.

Planning preliminaries Classical planners Learning modules Neural planning Challenges Applications
Survey structure

Interactive Taxonomy

Toggle between the major layers of the survey. Each tab shows the role of the component, the relevant neural methods, and the figure that best supports that part of the paper.

Deep Learning Modules

Which model family fits which planning role?

Neural planners differ less by name and more by what they represent: configuration samples, full trajectories, collision proxies, distance fields, or scene-conditioned policies.

Paper Map

An expandable tree of the reviewed papers, organized using the taxonomy from Section V of the survey. Search by title, model family, year, or planning role.

Loading reviewed-paper tree...

Challenges and Opportunities

01

Generalizability

Move from narrow training distributions to unseen robots, objects, obstacles, and task contexts.

02

Safety

Combine fast neural inference with verifiable constraints, collision guarantees, and runtime monitors.

03

Data and Simulation

Scale trajectory data while reducing dependence on expensive expert planners and brittle simulators.

04

Evaluation

Benchmark success, planning time, cost, smoothness, safety, and out-of-distribution robustness consistently.

05

Foundation Models

Bridge high-level generalist policies with low-level motion generation and physical feasibility.

06

Domain Deployment

Adapt neural planners to healthcare, re-manufacturing, and other settings where risk and constraints matter.

Autonomy stack

Manipulation autonomy stack

Deep learning modules

Core neural modules

Deep generative models

Generative model families

BibTeX

@article{soleymanzadeh2026towards,
  author  = {Soleymanzadeh, Davood and Lopez-Sanchez, Ivan and Su, Hao and Li, Yunzhu and Liang, Xiao and Zheng, Minghui},
  title   = {Towards Generalist Neural Motion Planners for Robotic Manipulators: Challenges and Opportunities},
  journal = {IEEE Transactions on Automation Science and Engineering},
  year    = {2026}
}