What mapping ocean floors taught me about data
10 July 2024
Potential’s CTO Bill Lamey spent two decades innovating sensors, software, and data in never-before-seen environments for IBM and Teledyne Caris. Here, he shares three key lessons that are accelerating our technological development.
“We know less about the bottom of our oceans than we do the surface of the moon,” says Bill Lamey. “Most of our seabeds have never been seen by humans.”
He should know. Before joining Potential, Bill was VP Engineering and Product at Teledyne Caris, the world leader in ocean floor mapping and the company where he spent a significant section of his 25-year engineering career.
(During that time, the company supported the coastal mapping of every developed nation around the globe; the data Bill and his teams produced underpin the charts that help ships navigate safely all over the world today.)
Bill has spent his entire career working at the forefront of emerging technologies and unmapped, unstructured and unseen environments. Now, as our CTO, Bill is applying 25 years of expertise and experience in sensors, data and processing to the unique challenges we’re helping our customers solve.
A brief history of Bill
Bill grew up in rural Canada, which means “a lifetime spent on motorized vehicles,” he laughs. “From ATVs and side-by-sides to motorcycles and snowmobiles, I’ve always enjoyed the calmness – and challenges – of navigating off-road.”
He graduated in 1999 from the University of New Brunswick with a Bachelor of Computer Science, specializing in hardware systems (digital design and programming). Soon after, he joined Caris, a spin-out from the same university, which had been formed to commercialize research into digitizing map information. “I was captivated by the idea that we were processing and modeling data of places that have never been seen, including actual shipwrecks!” says Bill.
Since then, Bill’s areas of expertise have grown to span processing and interpreting remote sensor data, large 3D modeling, parallel processing and computing, AI and cloud computing. He’s held a variety of technical, R&D and product roles, culminating in senior leadership positions at IBM and Caris (which was acquired by Teledyne during his tenure) and introducing the first AI-powered data-processing platform in the marine mapping industry.
Here, Bill explains how lessons he’s learned about data and computing from a quarter of a century of technological innovation are now underpinning our approach to reading unstructured environments.
1. Data is king – but it can be friend or foe
“Having access to a lot of data doesn’t always help. It's having the right data at the right time that matters.”
Mapping oceans requires many different sensors and data types, including dual GPS and RTK, IMUs (inertial measurement units), tides, speed of sound through the water and single-beam and multi-beam sonar.
“Using raw, time-based data – including from different ships – the software I helped build would create a vast, georeferenced point cloud,” Bill explains. “We built visualization tools to then create 3D models of the seafloor, and turn these into navigational charts.”
The data was collected using specially-equipped ships, which brought several challenges. “Ships are incredibly expensive to operate,” explains Bill. “Coming back to get more data or spending a lot of time on site weren’t options. We had to process data efficiently and robustly to allow ships to move to the next survey area.”
Quality control was critical. “I built a lot of tools over the years to help adjust and correct data that was not calibrated properly, and we developed algorithms to help automatically determine and test these in the field,” says Bill. “The safety of shipping all over the world relies on this data: getting it right was non-negotiable.”
When it comes to Potential’s Terrain Intelligence technology, data is a familiar theme. We’re building technology that can help vehicles recognize and interpret surface and terrain characteristics in off-road environments, which are unmapped and unstructured.
“We’re crunching through a huge amount of data, in real-time and computing onboard the vehicle,” says Bill. “Learning from my previous projects, we’ve built several tools at Potential that we can use in the lab or on the vehicle to make sure that we pull out just what we need to train and verify our AI models. These tools ensure our data is always properly calibrated. They’ve already reduced our training time by hundreds of hours, and we know we can trust the results.”
2. If the data doesn’t exist, do something about it
“We looked at libraries, at open-source, and at partner databases - and we realized we needed to do this ourselves.”
What we’re doing with Terrain Intelligence is such a new field that accessing good-quality, appropriate data for these environments has proved difficult. In many cases, it simply doesn’t exist. So we’ve been conducting our own data collection, sorting and processing with our own fleet of test vehicles, which includes motorcycles, side-by-sides, pick-up trucks and SUVs.
“We’ve designed 75% of Terrain Intelligence features to read from a single, industry-standard onboard car camera - but to ensure we have the best quality data and verification to train our AI models, we’re using a variety of sensors in testing,” Bill explains.
Across Potential’s fleet of test vehicles, we’re using mono and stereo cameras, radar, lidar, IMUs, and vehicle data such as steering wheel angle. We collect data every week at our off-road test track and a wide range of off-road locations, throughout the year and in different weather conditions. (We include on-road data collection, too.) The result is a unique library of detailed, accurate, and appropriate data that continues to grow.
“The challenge is the same as it’s always been throughout my career: finding innovative ways to capture and transmit only what’s important,” says Bill. “The good news for our customers is that Potential has been focused on data collection in off-road environments since 2018, which gives us quite the head start.”
3. You can do big things with tiny compute capacity
“Squeeze everything you can get out of computers. Never assume you’ll have unlimited compute resources.”
When Bill started out, cloud computing was in its infancy and AI was mostly theoretical. “The best workstations had one or two single-core CPUs and 512MB to 1GB of RAM,” he says. “Data collection rates were in the megabytes-per-minute range. Post processing was measured in days or weeks.”
While hardware and compute capability have advanced considerably in the intervening two decades, the need to do more with less remains a consistent thread in the challenges Bill has faced.
“Today we can collect multiple gigabytes of data per minute, but all that data still needs to be assessed, categorized, and processed in a timely way, and that requires ever more hardware – it’s a bit of a loop,” says Bill.
This raises a particular challenge when it comes to the automotive sector. “Cars are not built with supercomputers,” Bill notes. “Most of our automotive clients don’t have vast reserves of available, unused compute - they’re balancing many competing demands for capacity, and thinking about future upgrades too.”
Our software is designed to run on existing compute modules without additional hardware. This means we are working hard to ensure we minimize the footprint of our platform at every stage of development.
“Really, it’s the same principle as when I started out mapping seabeds back in the 1990s,” Bill points out. “It’s all about figuring out how to overcome huge obstacles with constrained resources. That’s where AI, married to inhouse technical know-how, is empowering us to achieve incredible leaps forward.”
In Bill’s next article, he’ll unpack our approach to AI, how we’re helping our customers adopt and adapt to emerging technologies, and why the right culture is vital in deep-tech start-ups.
Get in touch
Would you like to explore what Terrain Intelligence could do for you, or what you could do for Potential? Bill is keen to hear from you. Drop us a line at info@potentialmotors.com