Olivabot Vision

Olivabot Vision is an MIT feasibility project (MIT-haalbaarheidsprojecten Noord-Holland 2026). With support from the Province of Noord-Holland, we researched whether computer vision can be added to Gibbon Bot for more precise fruit monitoring and harvesting.

The study ran from 8 April 2026 to 31 July 2026.

This project is made possible with the support of the Province of Noord-Holland.

What we studied

The feasibility work focused on giving Gibbon Bot the ability to see, without compromising its tree-climbing design. We looked at camera integration, AI fruit recognition, robot control, on-board computing, communication in orchards, and outdoor robustness.

Findings

• Camera on the robot – A compact edge-AI camera module can be integrated into the existing architecture without upsetting the weight balance needed for climbing.
• Fruit recognition – Computer vision models (including CNNs and occlusion-aware segmentation) can detect fruit under changing light, shadows, reflections, and when more than half the fruit is hidden by leaves. Lightweight detectors suitable for an on-board camera were selected. Reaching high accuracy across crops still needs more training data and pipeline work.
• Vision and motion – Latency and real-time data between the vision pipeline and the robot’s control loops were analysed in simulation, including reinforcement learning in virtual environments. Linking camera output to actuators is feasible; safe field use still needs a further R&D phase.
• On-board AI (edge) – Image analysis can run locally on the robot, without a cloud connection. Lightweight network architectures were benchmarked on low-power embedded hardware. Computing is best split: the camera for recognition, local microcontrollers for each arm segment, and a central unit to coordinate vision and motion, within thermal and energy budgets.
• Communication in orchards – Offline operation is preferred for safety and cybersecurity. Low-bandwidth long-range radio is too slow for this data. Local Wi-Fi with on-robot storage and later sync to a more powerful computer is the practical setup. The robot stays lighter and can work independently.
• Outdoor robustness – The camera housing needs an IP65-level seal against dust and rain, using a robust transparent cover, while staying light enough and leaving room for heat management so the camera does not overheat.

Grower interest

From interviews with growers, interest splits. Some want analytics: fruit ripeness for harvest timing, and early signs of pests or disease. Others are more interested in the harvest itself. It depends on the farm.

Partners

Thanks to KC-Group for supporting our subsidy application and helping make this project possible, and to the Province of Noord-Holland for co-financing it.