Reinforcement Learning for the Full Strawberry Harvesting Process: Obstacle Separation, Detachment, and Placement

2026-07-16

AI Summary

This work addresses the complex contact dynamics in strawberry harvesting caused by severe occlusion and plant deformation by proposing an interaction-aware unified reinforcement learning framework that formulates obstacle separation, fruit detachment, and placement as a sequential decision-making task. The approach employs a hierarchical architecture integrating a high-level policy with low-level Cartesian impedance control, leveraging a shared interaction-aware policy to generate motion across phases and lightweight heuristic logic to coordinate task sequencing and gripper actions. Zero-shot sim-to-real transfer is achieved through feasibility-prioritized observation alignment and domain randomization. Experiments demonstrate a success rate of 89.7% in simulation and 82.0% on physical hardware, with average execution times ranging from 12.99 to 21.73 seconds across occlusion levels 1–5, confirming the system’s cross-platform efficacy and robustness.

Abstract

Severe occlusions and deformable plant structures introduce complex contact dynamics that challenge robotic strawberry harvesting. A policy-driven reinforcement learning (RL) framework with heuristic phase coordination was developed, in which obstacle separation, fruit detachment, and placement were formulated as a sequential decision-making task. A shared interaction-aware policy generated Cartesian motions across all task phases, while lightweight heuristic logic coordinated task progression and gripper events. A shared structured observation space was used to represent target, obstacle, end-effector, and task-context information. A hierarchical architecture combined the high-level policy with low-level Cartesian impedance control for compliant interaction. To support zero-shot sim-to-real transfer, feasibility-first observation alignment and domain randomization were adopted. The policy achieved success rates of 89.7% in simulation and 82.0% in real-world experiments. As the occlusion level increased from 1 to 5, the average execution time increased from 12.99 s to 21.73 s, reflecting greater interaction complexity. These results demonstrated effective transfer of interaction-aware harvesting behaviors to a structurally different robotic platform.

Problem

• strawberry harvesting
• occlusion
• deformable structures
• contact dynamics
• robotic manipulation

Innovation

• reinforcement learning
• sim-to-real transfer
• obstacle separation
• interaction-aware policy
• hierarchical control