Robotic harvesting demands precise end-effector positioning within constrained workspaces, particularly when navigating around delicate crops like hydroponic lettuce. This study introduces Path-Safe Approach Proximal Policy Optimization (PSA-PPO), a target-conditioned local reaching framework designed specifically for robotic harvesting of nutrient film technique (NFT) hydroponic lettuce in greenhouse environments. The method preserves the standard clipped PPO objective while incorporating harvesting-specific policy formulation elements: target-conditioned observations, corridor-aware spatial constraints, trajectory-history-based safety encoding, and task-defined success criteria for constrained robotic manipulation. The entire training pipeline runs within Isaac Lab, and the resulting policy transfers directly to a physical UR10e robotic arm without any real-world fine-tuning.
Extensive physical evaluation against 150 scheduled targets demonstrates that PSA-PPO achieves a 94.7% reaching rate, outperforming Cartesian+TOTG (86.0%), RRTConnect (88.0%), and RRT* (80.0%). Beyond reaching rate, the method reduces median RMS jerk by 39.8 to 73.0% relative to classical planners during the final approach segment, indicating smoother and more biologically plausible motion profiles. An ablation study over 200 simulation targets confirms that path-safety and side-corridor shaping are the principal drivers of harvesting-compatible performance. Strikingly, removing either term substantially degrades strict success rates even when terminal endpoint accuracy is preserved or improved, confirming that optimization toward the endpoint alone is an insufficient objective for constrained crop-proximal manipulation.
The authors further integrate PSA-PPO with an RGB-D vision module and physical end-effector, evaluating the complete greenhouse harvesting system across 144 lettuce harvesting attempts. The system achieves 94.4% vision-guided target reaching and 86.1% complete harvesting success. This end-to-end demonstration validates that the sim-to-real transfer not only works for reaching but also under full perception and actuation loops required for real-world harvesting tasks. The 86.1% success rate, while leaving room for improvement, represents a significant advancement for autonomous greenhouse robotics and suggests that the path-safety and corridor-aware design principles are effective for real-world deployment.