Adaptive Hybrid Optimization Learning-Based Accurate Motion Planning of Multi-Joint Arm.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Traditional motion planning struggles with dynamic, high-dimensional environments.
• Neural motion planning (NMP) offers a promising alternative.
• Training NMPs for high-accuracy tasks remains challenging.

Purpose of the Study:

• To develop an improved neural motion planning algorithm for robotic manipulators.
• To enhance efficiency and accuracy in dynamic and high-dimensional planning spaces.
• To address the difficulties in training neural networks for high-accuracy motion planning.

Main Methods:

• A hybrid approach combining Artificial Potential Field (APF) and reinforcement learning (RL) was proposed.
• The Soft Actor-Critic (SAC) algorithm was used to train the neural motion planner due to the high-dimensional, continuous action space.
• The method integrates wide-range obstacle avoidance with APF for precise position adjustments.

Main Results:

• The hybrid APF-RL method demonstrated a higher success rate in high-accuracy planning tasks compared to using APF or RL alone.
• Simulations validated the effectiveness across varying accuracy requirements.
• The learned neural network was successfully transferred to a real manipulator for dynamic obstacle avoidance.

Conclusions:

• The proposed hybrid APF-RL method significantly improves motion planning performance for robotic manipulators, especially in high-accuracy scenarios.
• This approach effectively combines the strengths of APF and RL for robust obstacle avoidance and precise control.
• The study confirms the practical feasibility of deploying learned NMPs on real-world robotic systems.