How Forterra Sped Up Model Iteration Cycles by 50% for Its Autonomous Ground Systems

Introducing Customer: Forterra

Forterra builds autonomous ground systems for defense and commercial missions, so its vehicles can go into harm's way and people don't have to. Its autonomous driving system, AutoDrive, runs on defense light vehicles and commercial yard trucks.

Those vehicles carry a mix of sensors, from visible light and infrared cameras to radar and LiDAR, depending on the mission. Forterra's perception team turns that sensor data into the models that let each vehicle see its surroundings and operate on its own.

Forterra builds by deploying rapidly: ship, watch the product in customers’ hands, and feed what it learns into the next version. That loop can only move as fast as the ground truth behind it.

The Challenge: Ground truth couldn’t keep pace

Forterra’s vehicles operate off-road, in unstructured terrain, which means a lot of unknown unknowns. The vehicles are always collecting data, and every unknown they uncover has to become accurate ground truth before the models can learn from it.

“We're not just running visible light cameras. We're running IR cameras. We're running radars and LiDARs and different technologies depending on the situations we're operating in. The accuracy and fidelity of those annotations are critical.” - Joe Putney, Chief Technology Officer, Forterra

Forterra was getting a couple of hundred sequences annotated a quarter, and turnaround took two weeks at the very minimum. Quality was poor, so there was a lot of back and forth with providers, and Forterra’s ML and perception engineers ended up reviewing sequences themselves, time they should have been spending on the models.

For a team that builds by deploying rapidly and iterating, that cycle time couldn’t keep pace with model iteration.

The Solution: A data pipeline, built around Forterra

Encord’s work with Forterra goes beyond annotation: the two teams have adapted architectures, data collection and processes around Forterra’s data. Forterra gets more ground truth at the quality bar it needs, and its ML engineers are back to improving models instead of reviewing sequences.

Every sensor in one pipeline: Encord handles 3D annotation on LiDAR alongside Forterra’s camera, infrared and radar data, all in one place.

A quality bar Forterra can count on: Encord’s annotators are trained on Forterra’s guidelines, so the data clears their quality bar consistently.

Engineers back on the models: Forterra’s team works in the same interface as the annotators, steering quality rather than producing it. “Everybody is using the Encord interface, participating in that process to make sure that the data that we’re pushing through is the best quality,” says Amanda Sgroi, Director of Perception in the AutoDrive product team.

Less integration work: The Encord SDK plugs straight into Forterra’s systems, so getting data in and out of annotation is faster, and takes work off the engineering team’s plate.

The Results: Faster model iteration and better-performing autonomous vehicles

Forterra now gets more ground truth, at higher quality, for less engineering effort.

• Faster cycles: Ground truth comes back faster and at 10x the volume, speeding up model iteration cycles by 50%.
• Better real-world performance: Forterra has seen a big difference in the performance of its systems over the last year, and has more confidence that its vehicles can operate and complete their missions successfully.
• Expanding use cases: Forterra is now in the next evolution of AutoDrive, deploying across commercial yard trucks and defense light vehicles, and integrating new payloads, sensors and effectors to automate more of each mission.