Using the Ramblr Data Engine, Harvest CROO teaches autonomous harvesters when to pick strawberries – at scale

Teaching a robot to make real-world decisions requires data – massive amounts of it – continuously updated and precisely labeled. Harvest CROO's 16 robotic harvesters scan every berry on every plant across constantly changing strawberry fields – making picking decisions in real time, every 2 to 4 days. Each harvester generates continuous multimodal video streams from multiple cameras. Without a structured way to manage, process, and annotate that data at scale, training smarter models becomes slow, expensive, and inconsistent.

• Continuous, high-volume video data generated across 16 autonomous harvesters

• Multiple camera angles per harvester producing large-scale multimodal datasets

• Unstructured data difficult to manage, organize, and annotate efficiently at scale

• Annotation bottlenecks slow model training and limit the ability to handle edge cases

• Dynamic field conditions requiring models that improve continuously, not just once

Video streams from every harvester and every camera angle are ingested and organized automatically, eliminating the manual overhead of handling large-scale multimodal datasets. Granular per-frame attributes and object relationships are annotated with precision – customized to the specific needs of Harvest CROO's picking models.

Active learning automates a significant portion of the annotation pipeline while preserving human oversight exactly where it matters: edge cases, accuracy verification, and ground-truth validation. The result is a continuously improving data engine that gets smarter with every harvest.

• Automatic ingestion and organization of large-scale, multi-view video datasets

• Multi-view support – synchronized data capture across multiple camera angles per harvester

• Granular per-frame annotation with customizable object relationships and attributes

• Active learning pipeline – automates repetitive annotation while flagging edge cases for human review

• Scalable foundation – continuously fed by real-world harvester data across every field cycle

With the Ramblr Data Engine, Harvest CROO replaced an annotation-heavy, bottleneck-prone pipeline with a scalable, automated training infrastructure – freeing their team to focus on model quality, not data management.

• Annotation overhead reduced from multiple annotators to one without sacrificing quality

• Faster training cycles through automated annotation and active learning

• More accurate picking decisions driven by higher-quality, continuously updated training data

• Edge cases systematically captured, verified, and fed back into the model

• Scalable data infrastructure that grows with every harvester, every field, every season