HarvestFlex: Strawberry Harvesting via Vision-Language-Action Policy Adaptation in the Wild

TL;DR

This paper introduces HarvestFlex, a vision-language-action policy system for real-world strawberry harvesting that achieves a 74% success rate with minimal real data, addressing occlusion and reflection challenges in unstructured greenhouse environments.

Contribution

First successful transfer of vision-language-action policies to real greenhouse strawberry harvesting using minimal real data and a multi-view RGB sensing system.

Findings

Achieved 74% success rate in real greenhouse harvesting

Reduced damage rate to 4.1% with fewer than four hours of real data

Demonstrated effective policy transfer without depth or geometric calibration

Abstract

This work presents the first study on transferring vision-language-action (VLA) policies to real greenhouse tabletop strawberry harvesting, a long-horizon, unstructured task challenged by occlusion and specular reflections. We built an end-to-end closed-loop system on the HarvestFlex platform using three-view RGB sensing (two fixed scene views plus a wrist-mounted view) and intentionally avoided depth clouds and explicit geometric calibration. We collected 3.71 h of VR teleoperated demonstrations (227 episodes) and fine-tuned pi_0, pi_0.5, and WALL-OSS with full fine-tuning and LoRA. Under a unified 50 trials real-greenhouse protocol and metrics spanning completion, pi_0.5 with full fine-tuning achieved success rate of 74.0% with 32.6 s/pick and damage rate of 4.1%. Asynchronous inference-control decoupling further improved performance over synchronous deployment. Results showed non-trivial closed-loop picking with fewer than four hours of real data, while remaining limited by close-range observability loss and contact-dynamics mismatch. A demonstration video is available at: https://youtu.be/bN8ZowZKPMI.