How Thoro Cut Model Deployment to One Week and Increased Labeling Speed by 50% with  Encord  

Introducing Customer: Thoro

Thoro is building an indoor autonomy stack for a broad range of mobile robots, with a primary focus on material handling in logistics and manufacturing environments, as well as industrial floor cleaning. For material handling, the company’s core vision is simple but ambitious: enable robots to move any pallet, anywhere.

Safety sits at the heart of everything Thoro builds, with the ongoing goal of achieving independent safety certification for all its robots, allowing them to operate reliably around both trained and untrained people.

Problem: A Demanding Deployment Environment and a Fragile Data Pipeline

Thoro’s most challenging deployment environments are large U.S. warehouses, up to 800,000 square feet, with no permanent racking. The only fixed landmarks are I-beams spaced roughly 30 feet apart, while the internal layout of pallets changes frequently.

This makes reliable navigation exceptionally difficult: robots must distinguish real aisles from temporary pallet storage areas with virtually no persistent spatial reference points.

Compounding this challenge was the team’s data and annotation infrastructure. Before Encord, Thoro managed all annotations manually using an open-source tool hosted on a local server. This setup created three major friction points:

First, because it was locally hosted, remote team members could not label data outside the office.
Second, dataset visibility was limited—the team could only see whether data was labeled, with no way to track progress or inspect pipeline flow.
Third, there was no automation: no auto-labeling, no automated data ingestion, and no mature pipeline to move data from robot to model. Everything had to be manually uploaded and downloaded.

As Chris Dunkers from Thoro put it:

Solution: A Cloud-Native Pipeline from Robot to Model

Thoro partnered with Encord to replace its fragmented, locally hosted setup with a fully cloud-based, automated data pipeline.

Data now flows from the robot into AWS, into Encord for labeling and quality review, and back out through Python training scripts before the updated model is deployed onto the robot.

The team primarily works with stereo image data (color images paired with aligned depth images), focusing on keypoint labeling and tracking for pallets across image sets.

Encord’s project and ontology tools provide a live view of data flow at every stage. One particularly valuable capability is the ability to add new ontology items and immediately reopen existing datasets to label only the new element, with full visibility into relabeling progress.

For model evaluation and experiment tracking, the team uses Weights & Biases alongside their Encord workflow.

Implementation: Operational Within a Week

Thoro’s onboarding experience was smooth from the start. The team ran a one-week trial and was fully operational by the end.

Connecting to their existing AWS infrastructure worked on the first attempt, and Encord’s support team provided fast, high-quality responses throughout.

Results: 50% Faster Labeling, One-Week Deployment Cycles, and New Customers Unlocked

Since adopting Encord, Thoro has significantly increased the speed at which it can respond to real-world model failures, a critical capability in live warehouse environments.

The team estimates they are at least 50% faster compared to earlier stages of the project, with each team member labeling or quality-controlling roughly twice as many images as before. The primary driver of this improvement has been the integrated auto-labeler, which performs a highly effective first pass on incoming data.

Model deployment timelines have shrunk to approximately one week: collect data on Monday, deploy an updated model by Friday. The team targets around 3,000 new images added within that window, a pace they are already achieving.

Beyond throughput, Encord’s pipeline visibility gave Thoro the operational insight needed to identify bottlenecks and optimize workflows. When one team member was labeling significantly faster than others, project insights helped surface and scale that workflow across the team.

The speed advantage has also unlocked new commercial opportunities. When Thoro deployed robots at a new customer site and encountered plastic-wrapped pallets, an edge case the existing model couldn’t handle, the team collected data, processed it through Encord, and deployed an updated model within three weeks.