Comparative Benchmark of Sampling-Based and DRL Motion Planning Methods for Industrial Robotic Arms

Area of Science:

• Robotics
• Artificial Intelligence
• Motion Planning

Background:

• Industrial robotic arms require efficient and reliable motion planning.
• Classical sampling-based planners (e.g., OMPL) face challenges in complex environments.
• Learning-based approaches, particularly deep reinforcement learning (DRL), show promise for improving motion planning.

Purpose of the Study:

• To compare the performance of OMPL planners against a DRL-based planner for industrial robotic arm motion planning.
• To evaluate planning time, success rate, and trajectory smoothness.
• To investigate the potential of DRL for real-time, high-throughput applications.

Main Methods:

• A UR3e robot with an RG2 gripper was used.
• Over 100,000 collision-free trajectories were generated using OMPL and MoveIt.
• A DRL agent was trained using curriculum learning and expert demonstrations (Soft Actor-Critic).
• Time-Optimal Parameterization using TOPPRA ensured dynamic feasibility.

Main Results:

• The DRL planner achieved higher success rates and significantly lower planning times compared to OMPL.
• DRL-generated trajectories were more compact and deterministic.
• Classical planners showed better zero-shot adaptability and environmental generality.

Conclusions:

• DRL-based motion planning offers significant advantages for industrial robotic arms, especially for real-time applications.
• Hybrid architectures combining DRL and classical planners could leverage the strengths of both.
• This study provides practical insights into planning paradigm trade-offs.