PITNet: Physics-Informed Trajectory Network for Smooth 7-DoF Robotic Arm Trajectories in Dynamic Environments

This paper introduces PITNet (Physics-Informed Trajectory Network), a novel framework for trajectory generation in 7-degrees-of-freedom (7-DoF) robotic manipulators operating in dynamic environments. PITNet developed by integrating physics-informed learning, curriculum-based training, and a multi-objective composite loss function. This approach uniquely combines dynamic modeling, obstacle-aware navigation, and supervised learning to overcome the limitations of traditional methods. In addition, PITNet simultaneously addresses positional accuracy, trajectory smoothness, energy efficiency, and obstacle avoidance. PITNet outperforms Quintic Polynomial Method (QPM) and Model Predictive Control (MPC) in five configurations. It achieves a 12.9% reduction in RMSE, a 99.5% decrease in jerk compared to MPC, and a 35.9% improvement in the minimum obstacle distance, with statistical significance in all metrics (p < 0.05). Although the QPM produces the lowest jerk (0.000319 rad/s ), but its RMSE is 12.9% higher than PITNet, and it clears obstacles 1.3% less effectively than PINet. Additionally, its computation time is much slower (1.03 s), which reduces its suitability for dynamic environments when compared to PITNet. MPC, though adaptable, suffers from excessive computational latency (48.07 s). In contrast, PITNet operates at an inference speed (0.0103 s) which is 99.0% faster than QPM and 99.98% faster than MPC. Moreover, the training loss for PINet converges from 1.2 to 0.6575 over 500 epochs, validating the effectiveness of the curriculum learning strategy. PITNet's physics-driven design advances multi-objective trajectory planning, with a strong potential for real-time adaptation and energy optimization in dynamic robotic environments.