Sim-to-Real Reinforcement Learning for Path-Safe Target Reaching in Robotic Harvesting of Hydroponic Lettuce

The 30-second take

• What: PSA-PPO is a target-conditioned local reaching policy with corridor-aware constraints and trajectory-history safety, trained in Isaac Lab and transferred to a UR10e with no real-world fine-tuning; it beats classical planners on a 150-target bench and, with RGB-D plus an end-effector, completes 86.1% of 144 lettuce harvests.
• Abundance angle: today, gentle crop-proximal manipulation is scarce skilled labor in constrained greenhouses. Sim-to-real path-safe reaching is a step toward cheaper default harvest robots if safety and success hold across farms (mid-horizon: reliability, safety, and unit economics decide defaults).
• Who should care: Agricultural-robotics groups, greenhouse operators on NFT lettuce, motion-planning teams stuck on jerk and corridors, and sim-to-real RL practitioners with a UR-class arm.

What the paper actually did

Robotic harvesting needs precise end-effector placement in a constrained workspace. The authors introduce Path-Safe Approach Proximal Policy Optimization (PSA-PPO), a target-conditioned local reaching framework for nutrient film technique (NFT) hydroponic lettuce in greenhouses. They keep the standard clipped PPO objective but add a harvesting-oriented formulation: target-conditioned observations, corridor-aware spatial constraints, trajectory-history-based safety encoding, and task-specific success criteria.

PSA-PPO was trained entirely in Isaac Lab and transferred to a physical UR10e with no real-world policy fine-tuning. On a controlled physical benchmark of 150 scheduled targets, it reached 94.7%, ahead of Cartesian+TOTG (86.0%), RRTConnect (88.0%), and RRT* (80.0%), while cutting median RMS jerk 39.8–73.0% versus those classical planners on the final approach. A formulation ablation over 200 simulation targets finds path-safety and side-corridor shaping as the main drivers of harvesting-compatible success: dropping either term hurts strict success even when terminal endpoint accuracy holds or improves — so endpoint-only optimization is not enough near crops.

Integrated with an RGB-D vision module and an end-effector as a full greenhouse system, it recorded 94.4% vision-guided target reaching and 86.1% complete harvesting success across 144 lettuce attempts.

What makes this disruptive

The scarce capability is a reaching policy that is not only accurate at the leaf but path-safe and smooth in a cluttered NFT channel. Zero-shot transfer from Isaac Lab to a UR10e, with 94.7% physical reaching and lower jerk than RRT-family planners, is a strong sim-to-real harvest result as reported.

The ablation is the conceptual sting: optimize only the endpoint and you can look accurate while failing the harvest-compatible criterion. Path-safety and side corridors are not garnish.

86.1% complete harvest on 144 tries is still a research greenhouse campaign, not a season-long commercial SLA. Classical planner names and percentages are the right comparison set for reaching, not for every agronomic failure mode.

Why it matters (outside the lab)

Abundance lens: picking delicate produce in tight hydroponic layouts is skilled, repetitive labor. If path-safe reaching becomes a default software layer on ordinary arms, more of that harvest capacity can be shared automation rather than scarce crews — after reliability and crop damage metrics hold.

Near-term: greenhouse robotics teams can treat corridor-aware RL as a reaching stack, not a full farm OS. Medium-term: crop variety, lighting, sanitation, and safety certification decide defaults. No harvest-robot launch year.

Unit economics and bruise rates remain the gate.

Limitations & open questions

Physical reaching is 150 scheduled targets; harvest is 144 attempts in the authors’ greenhouse setup. No real-world fine-tuning is a feature of this study, not a proof that every new lettuce line will transfer. Ablation is simulation (200 targets).

Vision, cutter, and plant-variability failures are only summarized as complete-harvest success (86.1%). Preprint ≠ commercial harvester. Abundance is not automatic: a reaching policy does not staff-out a greenhouse on a schedule.