How Ambi Trained Sorting Robots to Achieve 99.9% Sort Accuracy

About Ambi Robotics

Ambi Robotics is on a mission to make ecommerce logistics faster, more accurate, and more scalable. Their flagship platform, AmbiOS, is a deployment layer for physical AI, connecting robot brains and bodies with critical infrastructure for safety, support, and operations. It is a software layer that enables robots to pick, pack, sort, and stack at industrial speed by learning from both simulated environments and live deployment data.

At the heart of their approach is a proprietary technique called Sim2Real AI: robots are trained in virtual simulations, and those learned behaviours transfer seamlessly to real-world operations. The result is faster deployment and higher reliability, which is a critical advantage when handling millions of parcels in commercial environments. AmbiOS continuously learns from over 250,000 hours of real-world data, and the company's AI Skill Suite gives enterprise customers access to production-proven capabilities including item intelligence, inspection, dexterous picking, and precision placement.

The Challenge: Picking accuracy depends on capturing real-world variety

Ambi Robotics is already in full commercial production, with its robots sorting ~150 million+ items across customer sites. Their suction-based picking systems have to handle whatever comes its way in different shapes and sizes. As their volume is high and their objects vary, edge cases arise often. 

That makes training data the limiting factor on picking accuracy. There is no large, pre-existing set of labelled parcel-picking data; every datapoint is collected from robots working in live environments. For models to generalise, annotations have to capture the full range of what the robots see, and the edge cases matter most. Ambi's needs also go beyond object detection: the team annotates the text on parcel labels and stickers to improve how reliably their systems can read it.

Ambi previously worked with another annotation platform, but it wasn't built for the precision they needed. An annotator might correctly draw a polygon around a parcel but leave out its category. This across thousands of annotations fed back into the model can affect their performance when it comes to real world variability.

"The annotator annotates an object with a polygon but forgets to mark its type or category. Those errors were getting through." - Qi, ML Engineer, Ambi Robotics

The Solution: Annotation infrastructure built for precision

Ambi Robotics switched to Encord to bring structure, validation, and predictability to their annotation pipeline. The impact showed up immediately in the quality controls they were able to enforce and in the reliability of data flowing into model training. The team was able to get consistent, on-time delivery of all annotations. And those annotations were high quality across all tasks.

Validation rules that prevent incomplete submissions: Encord's required field validation ensures annotation quality by enforcing that all mandatory objects, classifications, and attributes are completed before submission; validated across ontology definitions, custom layouts, and individual tiles in multi-tile data groups. Errors are caught early rather than when the model has already been deployed.

Rigorous review workflows: Encord's workflow has multi-stage review, ensuring errors don't slip through. This ensures that the data pipeline is both consistent and complete. 

A developer-friendly SDK: Ambi's engineering team built Encord's SDK directly into their automated pipelines, so data moves from deployed robots into annotation and back into training with minimal manual handling. For a fast-growing team, that means engineering time goes into model development rather than data infrastructure.

Full visibility into annotation progress. Encord's analytics give Ambi a live view of their annotation workforce, including throughput, task progress, and quality by annotator and review stage. ML engineers can see exactly where each batch stands, which lets them plan training runs and iteration cycles with confidence.

"Encord's workflow system lets us enforce more rigorous review, so annotations don't slip through with errors. The validation rules are especially useful. We can make fields required so an annotator cannot submit a task with a missing label." - Qi, ML Engineer, Ambi Robotics

The Results

The improvements in annotation consistency and completeness have fed directly into Ambi’s model quality. allowing them to push toward new performance benchmarks and expand the capabilities of their robotic systems.

These include:

• 99.9% sorting accuracy: a new performance ceiling, enabled by cleaner training data
• Consistent throughput: annotation volume stays steady so the ML team can plan training runs around a reliable flow of labelled data rather than waiting on bottlenecks
• Expanding into new use cases: For example, their sorting algorithm can now work on a variety of containers and items; their AmbiStack stacking robot, expanded container type sorting, and higher-reliability operations across a wider variety of items and conditions. 

"It's where raw images become labelled training data; with workflow management and quality control."
- Qi, ML Engineer, Ambi Robotics