Kiwibot’s autonomous delivery fleet operates across complex pedestrian environments — dynamic settings where sidewalks, roads, and seasonal conditions are constantly changing. Each robot continuously captures images from multiple cameras, uploading them to the cloud — generating millions of images for potential use in training their segmentation models.

However, collecting data wasn’t the hard part — the real challenge was deciding which images were actually worth labeling.

Their models needed to adapt quickly, but labeling is expensive, and selecting the right data to annotate was tedious, manual, and error-prone. 

Before Lightly, Kiwibot relied on GPS data and random sampling to select images for training. The process was time-consuming, repetitive, and often surfaced redundant or uninformative scenes - especially in environments like long, uniform sidewalks or intersections with no clear boundaries.

Kiwibot replaced this manual selection process with a programmatic pipeline built around Lightly Worker, Lightly’s code-first data selection engine. 

The team integrated Lightly directly into their data pipeline, enabling automated selection of the most diverse and informative samples from millions of robot-captured images. Their system now:

• Combines model uncertainty signals, camera metadata, and Lightly’s diversity-based sampling to flag edge-case images
  
• Runs continuously, fully integrated into the training workflow — no human-in-the-loop selection required
  
• Outputs only high-value frames for annotation, reducing label waste and improving iteration speed

Lightly’s embedding visualizations also gave early confidence that selection was working well, while Kiwibot appreciated the hands-on support and constant updates from the Lightly team.

“Lightly isn’t just a tool - the team actively shares updates, model suggestions, and helps us move faster. It’s not something you get from most companies.” - Carlos Alvarez

Kiwibot selected Lightly because it directly addressed a critical problem: the inability to consistently identify and label edge cases that matter. With over a million images coming in from their robot fleet, randomly sampling data wasn’t cutting it — it often resulted in redundant, uninformative samples that did little to improve model performance.

Lightly enabled them to shift from reactive to targeted retraining by automatically surfacing the types of data that cause real-world failures in production.

Common edge cases include:

• Sidewalks merging into roads without clear visual separation, confusing the segmentation model
  
• Snow-covered surfaces, which dramatically change visual features and scene structure
  
• Puddles and potholes, often misclassified due to lack of representation in the training set

These scenarios were nearly impossible to catch with GPS-based or heuristic sampling. With Lightly’s selection pipeline plugged into their training loop, the team can now systematically surface these failures — leading to faster, more focused retraining cycles.