Deep Reinforcement Learning Based Trajectory Planning Under Uncertain Constraints.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Theory

Background:

• Traditional trajectory planning struggles with high-dimensional, uncertain environments.
• Deep Reinforcement Learning (DRL) offers autonomous learning capabilities for complex robotic tasks.
• Safe human-robot coexistence requires robust collision-avoidance in trajectory planning.

Purpose of the Study:

• To present state-of-the-art DRL-based collision-avoidance trajectory planning for uncertain environments.
• To compare model-free, policy gradient-based DRL algorithms (SAC and DDPG) for robot manipulator trajectory planning.
• To evaluate the effectiveness of DRL in enabling autonomous, safe, and accurate trajectory planning.

Main Methods:

• Simulation of a 7-DOF Panda robot manipulator using the PyBullet physics engine.
• Adaptation and comparison of Soft Actor-Critic (SAC) and Deep Deterministic Policy Gradient (DDPG) algorithms.
• Evaluation metrics included reward, loss, safe rate, and accuracy over 5,000 training episodes.

Main Results:

• DRL algorithms demonstrated effective trajectory planning in simulated uncertain environments.
• Both SAC and DDPG frameworks were adapted for high-dimensional continuous state-action spaces.
• The proposed DRL approach achieved zero collisions after extensive training.

Conclusions:

• DRL algorithms are highly effective for collision-avoidance trajectory planning in uncertain environments.
• Autonomous learning via DRL surpasses traditional methods in complexity and safety.
• The study validates the potential of DRL for safe human-robot interaction scenarios.