Abstract:To meet the need for real-time thrust-command generation during the powered-descent phase of planetary landers under multiple constraints—and to address the low-latency requirement caused by limited onboard computing resources—this paper proposes an intelligent control method based on physics-constrained imitation learning. The method takes a sliding time window of powered-descent states as input, enforces data consistency via feature normalization and inverse normalization of thrust, and uses a one-dimensional convolutional network to extract short-horizon temporal dynamics and output thrust commands. To ensure the physical feasibility of thrust predictions and compliance with trajectory constraints, a physics-constrained loss is introduced during training: differentiable dynamics explicitly couple the predicted thrust to trajectory evolution, while glide-slope corridor constraints and thrust-magnitude feasibility bounds are encoded as soft penalties, thereby embedding geometric feasibility and actuator realizability directly into the learning process. Compared with behavior cloning based on pure data regression, the proposed strategy effectively suppresses infeasible action outputs induced by out-of-distribution samples, improving policy robustness and engineering practicality. Results show that the method significantly reduces online computational burden while maintaining soft-landing accuracy and robustness, making it suitable for real-time deployment on compute-constrained onboard platforms.