Terrain Intelligence 2024 release shrinks AI model by 75% and halves inference time
05 November 2024
This week, we released our 2024 model for Terrain Intelligence, which cuts inference time and initialization time in half, and model size by 75% - and all without losing any accuracy. A smaller, faster model means easier adoption for our automotive and powersports customers.
After months of work, we’ve just released major updates to Terrain Intelligence, our core technology that uses a single forward-facing camera to deliver real-time insights about off-road surfaces to help drivers better navigate extreme environments. We’ll unveil all the new features in the coming weeks.
Following extensive optimization work, our new 2024 model is just 25% the size of last year’s model.
Inference time (how long it takes an artificial intelligence model to ingest information and make decisions) has been cut to 45% of last year’s model - in other words, its speed has more than doubled.
And initialization time (how long it takes for the artificial intelligence to get ready for work) has been reduced by 53%. Importantly, our model maintains its accuracy even following this extensive optimization.
Terrain Intelligence 2024 model cuts inference speed by 55%, initialization time by 53%, and model size by 75% - and all without losing any accuracy.
Benefits for powersports and automotive customers
“A smaller, faster model is vital for all our customers,” explains Sam Poirier, our CEO.
“For powersports customers who are just beginning the journey into onboard cameras and embedded compute, it means they can specify lightweight, affordable hardware systems and still unlock all the benefits of Terrain Intelligence.
“Our automotive customers have far more comprehensive onboard compute capacity, but they are also balancing many more demands for that space. A smaller footprint makes adoption far easier.”
Terrain Intelligence unlocks insights into surface characteristics - such as roughness, friction, incline changes and hazards that would affect suspension travel - and provides feedback that allows vehicles to self-select the most appropriate drive, suspension and traction settings for the path ahead. The platform has been built particularly with unstructured, off-road environments in mind, but also solves edge cases on the road.
For powersports customers, Terrain Intelligence can enhance performance and safety. That means a more accessible activity for newcomers and youngsters, and new horizons of fun and capability for experienced riders.
For automotive customers, Terrain Intelligence can add off-road credentials to vehicles that stay on the road most of the time. But even on the road, our platform can boost comfort, safety and performance in ways that non-novel, reactive technologies (that rely on wheel sensors and IMU data, for example) cannot deliver on their own.
“A smaller footprint makes adoption far easier for both automotive and powersports customers.”
Dramatic improvements underscore our AI-focused approach and talent
“These impressive leaps forward in Terrain Intelligence are a testament to the talent of our people and our use of AI,” says Bill Lamey, our CTO.
This summer, we’ve added several machine learning experts to our team, who are helping to turbocharge our development. We’re also running several different test vehicles, including a side-by-side, a motorcycle, a pick-up truck, and our own custom-built, all-electric, quad-motor test platform, P0.
Thanks to the power of AI, Terrain Intelligence doesn’t need maps, a cloud connection or sophisticated sensors like lidar. It works in real time, even at high speed, and the models are so agile that they can run onboard small vehicles like motorcycles and side-by-sides.
We will release more information about new features in the coming weeks.
To stay up to date with our exciting developments, sign up to our newsletter and follow us on LinkedIn.