Carv 2 Deep Dive: How an AI Coach Trained on 500 Million Turns Reads Your Skiing

Photo from https://getcarv.com/blog/introducing-carv-2 of the Motion Metric's Carv 2 product

I've spent a lot of time obsessing over how machines sense the physical world, so a device that claims to feel my skiing through two clip-on sensors was always going to end up on this blog. Carv 2 is the second generation of the digital ski coach from London-based Motion Metrics, and it represents a fascinating engineering pivot: they threw away most of their hardware and replaced it with data. Let's dig into how that actually works.

From 48 Pressure Sensors to Zero

The original Carv was a marvel of instrumentation: a thin insert under each boot liner packed with dozens of pressure sensors, plus a motion unit, all reading how your foot loaded the boot through every phase of a turn. It worked, but anyone who's designed hardware knows the failure modes were baked in. Thin flexible circuits living inside a ski boot, flexed thousands of times a day at -10°C, connected by a cable to an external battery pack. Fitting them was fiddly, and they wore out.

Carv 2 deletes the pressure inserts entirely. Each unit is now a self-contained pod that clips to the outside of your boot cuff in seconds, built around a six-axis inertial measurement unit, an accelerometer paired with a gyroscope, sampling your boot's motion in three-dimensional space. No inserts, no cables, no boot surgery.

Here's the part I love: how do you recover the pressure information you just threw away? With data. Carv had logged hundreds of millions of real turns from the first-generation system, where motion data and pressure data existed side by side. Train a model on that corpus and it learns the correlation, so that from motion alone it can infer what the pressure profile must have been. The company says the new system, trained on over 500 million turns and cross-referenced against thousands of hours of video and instructor input, actually scores 6% more accurately than the sensor-stuffed original. It's a textbook example of a modern engineering trade: replace fragile physical sensing with a learned model and a simpler, more robust device. Your phone camera does the same thing every time computational photography outperforms a bigger lens.

What Ski:IQ Actually Measures

Every run gets distilled into a Ski:IQ score, a number roughly from 60 (survival wedge) to 160+ (instructor-grade carving). Under the hood, the system evaluates dozens of metrics grouped into balance, edging, rotary control, and pressure management, which anyone who's taken a PSIA lesson will recognize as the standard skills framework. The difference is resolution: an instructor watches you make ten turns from behind; Carv measures every single turn, all day, with the same criteria it learned from experts like Ted Ligety and hundreds of instructors who helped label what "good" looks like.

The current algorithm generation, Ski:IQ "Nevado," fixed my biggest gripe with earlier versions: terrain awareness. The old scoring implicitly assumed groomers, so a day in moguls or powder tanked your number even if you skied it well. Nevado detects the snow surface and terrain type and scores you against the skills that actually matter there. Absorb and extend well in bumps and you're rewarded for it, rather than penalized for not laying trenches.

Coaching in Your Ear, When You Want It

The app offers three modes. Train gives real-time audio through your headphones, including drills that react to every turn — get forward on the outside ski and you hear it instantly. Learn waits until the chairlift, then delivers a short personalized tip based on the run you just skied. Track shuts up entirely and just logs data for later.

That immediacy matters more than it sounds. In motor learning, feedback delay is everything; being told on the chairlift that your hips were back three runs ago is nearly useless, but hearing it mid-turn closes the loop while the sensation is still in your body. It's the same philosophy I wrote about in my Kaizen post: small, immediate, measurable corrections compound into real change. Carv just automates the measurement half of that loop.

By Carv's own numbers, over 56,000 skiers have now logged more than a billion turns through the platform, and the company claims 95% of members measurably improve. Self-reported platform stats deserve a grain of salt, but the independent reviews this past season (GearJunkie, TechRadar, T3, Advnture) broadly agree: for intermediates through experts, the feedback is accurate and the improvement is real.

The Catch: It's a Subscription

The business model is where opinions split. The sensors themselves are cheap (US retailers list the Carv 2 hardware around $249, and the bundled unlimited membership deals bring hardware cost to effectively zero), but the coaching lives behind a membership: roughly $199–249 per year for unlimited skiing, or a cheaper 6-day pass for occasional skiers. Let your membership lapse and the sensors keep tracking runs and Ski:IQ, but real-time coaching and detailed metrics go dormant.

I get the grumbling about subscription creep, but the engineering economics make sense here. The product isn't the pod on your boot, it's the continuously retrained model behind it, and that model improves every season precisely because members keep feeding it turns. Compare it to human instruction, where a single full-day private lesson at a major US resort now runs $900+, and a season of unlimited AI coaching for a fifth of that price looks like a bargain. Compare it to a $300 pair of goggles, and it's just another line item.

Who Should Actually Buy One

If you're a wedge-turning beginner, save your money and take a human lesson; Carv's own materials say it works best once you're skiing parallel. If you're a strong intermediate stuck on a plateau, this is the most cost-effective coaching tool I've seen, and the data agrees with me. And if you're an expert, the appeal is different: it's a measurement instrument. Like tracking your lines in Strava using FATMAP technology, having objective numbers on something you thought you understood by feel is occasionally humbling and always interesting.

Two sensors, one learned model, half a billion turns of training data. The interesting part isn't that an AI can coach skiing — it's that deleting hardware and substituting data made the product better. That's a pattern worth watching across all of ski tech.