AI Form Correction for Hyrox: 2026 Feasibility Assessment

“Many AI companies have promised real-time form correction. None have delivered.” — user, May 2026. This document is a technical reality-check, not a pitch deck.

1. The graveyard: who tried and what actually happened

Lululemon Mirror — the biggest cautionary tale

Lululemon paid ~$500M for Mirror in 2020, took a $442.7M post-tax impairment in Q4 2023, discontinued sales by Dec 2023, and laid off the 119-person NY Studio team in early 2024. Mirror never shipped meaningful AI form correction — it was streaming classes on a reflective display with optional heart-rate integration. Reviews explicitly cite this gap: “The lack of tactile or AI-driven form correction means users performing complex or heavily loaded movements risk reinforcing poor technique.” (Yahoo Finance, Retail Dive, BodyFreedom review)

Tonal Smart View — the most-shipped, most-criticized form coach

Tonal’s Smart View uses the phone’s front camera, claims “500 data points” between cable load cells and computer vision, and gives form cues on a handful of pre-tagged movements. Real user feedback is split:

Marketing-aligned reviewers like it: “like a personal trainer that took no BS” (GearJunkie)
Long-term users are blunter: “I’ve never seen [the camera] come up except possibly for some movements” (Ian’s blog)
Reliability complaints recur: app crashes mid-set, Smart View requires re-launching every few exercises (Trustpilot)

The honest read: Smart View works on a curated allow-list of movements with the user pinned to a single camera angle on Tonal’s rig, and even then is treated as a novelty rather than a daily driver.

Tempo Studio — alive, AI claims credible but unverified

Tempo is operationally alive in 2026 (no bankruptcy — that was the unrelated Tempo Automation PCB company). Their 3D depth-camera-based system is technically the most credible at-home form-coaching hardware shipped, but it remains a niche $2K+ device and there’s no independent third-party validation of form-feedback accuracy. (Tracxn)

HomeCourt (NEX Team) — the survivor, but narrow

Still alive in 2026 with an NBA equity partnership and 170-country user base. Crucially, HomeCourt succeeded by narrowing scope: it tracks shot makes/misses, release time, vertical jump, and dribble speed — not “form correction.” It counts and measures, it doesn’t coach. (NBA press release, HomeCourt.ai). This is the most important pattern in the survivor list.

Asensei — alive as B2B SaaS, not B2C

Asensei pivoted years ago from a consumer product to a B2B coaching-intelligence API. Customers are connected-fitness hardware OEMs (Centr, Vertimax, PowerBlock, Litesport). Survival strategy: don’t try to be the product, sell the picks-and-shovels. (Asensei.ai, Athletech News)

Vi.AI (LifeBeam) — dead

Vi running coach app shut down March 2022 after burning through hardware sales. Headphones still function as headphones but the coach is gone. (Trail and Kale)

Onyx — acqui-killed

Acquired by Cure.fit in Jan 2021. The product still exists nominally but reviews are damning: “calibration is beyond TERRIBLE… only about 1/3 of reps being counted” (JustUseApp reviews). The original mission of AI form correction was effectively abandoned.

Freeletics, PIVOT Yoga, Fitness Coach AI

Freeletics in 2026 explicitly does NOT correct form — they walked it back: “As it currently stands, the AI will not correct your form.” They show video demos instead (Dr. Muscle review). PIVOT Yoga uses sensor-embedded clothing instead of vision — a different (more expensive, less scalable) bet. Onyx-class pure-vision indie apps continue to launch and quietly underperform.

Pattern from the graveyard

Every consumer-vision form-coach that aimed broad has either died, pivoted to B2B, or quietly walked back the “AI corrects you” claim. The survivors narrowed scope to measurement (counting, timing) rather than correction (coaching).

2. Where the tech actually is, May 2026

2D pose estimation: solved for fitness reps

MediaPipe BlazePose and MoveNet are mature and effectively free. MoveNet Lightning runs at 192×256 in <7ms on a 2022-era phone; Thunder runs at 256×256 in ~20ms. Validation studies show Pearson correlation 0.91 for upper-limb, 0.80 for lower-limb vs gold-standard mocap, and >99% rep-counting accuracy on common bodyweight exercises (squats, push-ups, jumping jacks, sit-ups, pull-ups). (Roboflow, PMC pose review, ACM action counting)

3D pose from a single camera: meaningfully improved but still error-prone

The 2025 AthletePose3D benchmark fine-tuned SOTA monocular 3D models on 12 athletic actions and dropped MPJPE from 214mm to 65mm — a 69% gain, but 65mm is still ~2.5 inches of joint-position uncertainty, which is the difference between “knee tracks over toe” and “knee caves” in a squat. (AthletePose3D arXiv)

Apple’s VNDetectHumanBodyPose3DRequest (iOS 17+) returns 17 joints in meters relative to camera, uses depth data when available, doesn’t require LiDAR but is more accurate with it. (Apple docs). Vision Pro full-body tracking is still not shipped — Apple cut it years ago because engineers “couldn’t make it reliable enough” and the team is still trying. (UploadVR)

Multimodal LLM video reasoning: the genuine 2024–2026 unlock

Gemini 2.5 Pro/Flash: native video input up to 3h (or 6h at low resolution), priced at $0.30/$2.50 per M input/output tokens for Flash, $0.10/$0.40 for Flash-Lite. At 1 FPS sampling, each second of video = 258 tokens, so a 30-second exercise clip = ~7,740 tokens ≈ $0.003 to analyze with Flash, ~$0.0008 with Flash-Lite. (Gemini API docs, PricePerToken)

A real working example is published: ChaosFit streams webcam frames at 1 FPS to Gemini Live and gets conversational form feedback. (Medium writeup). Latency is 1–3 seconds round-trip — fine for between-rep coaching, not fine for in-rep correction.

GPT-4o Realtime / GPT-4.1: speech-to-speech under 320ms, state-of-the-art on Video-MME long-context (72.0%, no subtitles). Video reasoning is competitive with Gemini. (OpenAI, GPT-4.1)

Open source: Qwen2.5-VL (Jan 2025) handles 1h+ video with dynamic FPS sampling and is the most credible self-hostable option, but you’ll spend $1K+/mo on GPU to run it for serious load. (Qwen2.5-VL paper)

On-device fitness AI

iPhone 16/17 Pro Neural Engine can run MediaPipe Pose + custom rep-counting + simple rule-based form heuristics in real time with negligible battery cost. What it cannot do on-device in 2026 is run a true multimodal LLM that reasons about form qualitatively. Apple Intelligence’s local model is text-only for fitness purposes.

3. The 8 Hyrox movements: per-station feasibility

#	Station	Detectability	Rep count	Form-error flagging	Main failure mode
1	SkiErg 1km	Hard	Hard	Medium	User is on equipment, occluded; phone has to see arm extension cleanly
2	Sled Push 50m	Medium	N/A (distance)	Easy: hip extension, head position	Camera follow: athlete moves through space
3	Sled Pull 50m	Medium	N/A	Medium: hip hinge, rope grip	Same camera-follow problem
4	Burpee Broad Jumps 80m	Easy	Easy	Easy: chest-to-floor, plank, jump distance	Multi-phase motion is well-studied; this is the prime candidate
5	Rowing 1km	Hard	Medium	Medium: catch/finish position, back angle	Equipment occludes hips, athlete is seated
6	Farmer’s Carry 200m	Medium	N/A	Easy: shoulder shrug, trunk lean	Camera-follow, lighting variance
7	Sandbag Lunges 100m	Medium-Easy	Easy	Easy: knee tracking, torso upright, depth	Excellent fit — alternating, repetitive, fixed plane
8	Wall Balls 100 reps	Easy	Easy	Easy: squat depth, ball-to-target, full extension	Static position, lateral view, prime candidate

The 3 you’d start with: Wall Balls, Burpee Broad Jumps, Sandbag Lunges. All three are stationary-ish, repetitive, single-plane, and have well-defined form failures that map cleanly to joint-angle thresholds. The 5 you’d not start with all involve either equipment occlusion (SkiErg, Row), continuous locomotion (sleds, carry), or both.

4. The real-time vs between-set tradeoff

Mode	Tech latency budget	What’s possible in May 2026
A. Post-workout review	Minutes	Solved. Gemini 2.5 Flash + 30-second clip + structured prompt = $0.003 and a credible coach-quality writeup. This is what to ship first.
B. Between-set feedback	2–10 seconds	Workable. MediaPipe rep counter + local rules engine flags errors per rep, summarizes after the set in voice. No LLM needed for the core; LLM only for friendly verbalization.
C. In-rep real-time correction	<300ms	Not viable for qualitative coaching. Local rule-based (“your knee just caved”) tone-cue: yes, with caveats. LLM-based “you’re losing tension in your core”: no — round-trip latency kills it.

The deal-breakers for C are not compute — they’re (1) athlete is moving fast and won’t react to a verbal cue mid-rep, (2) pose jitter on a single phone camera produces false positives that destroy trust within 5 minutes, and (3) the athlete usually can’t see/hear feedback while their head is down or they’re under load.

5. AR/VR overlay reality

The “ghost overlay of correct form” idea is not shipping in any consumer product in May 2026.

Apple Vision Pro: Still no full-body tracking. Fitness apps that exist (Supernatural, ported FitXR, Tripp) are immersive cardio/meditation, not form correction. “Zero haptic correction. It can show you’re leaning too far forward in a lunge — but cannot feel the imbalance.” (Mixed News, UploadVR)
Snap Spectacles: Hand tracking only, no body tracking sufficient for full-movement coaching.
ARKit ghost overlay on iPhone: Technically possible (Apple’s ARCoachingOverlayView plus 3D body pose), but no shipped fitness app has made this useful. The fundamental problem: the athlete is looking at themselves doing the movement, not at the screen.

Honest verdict: AR ghost-overlay for fitness is a 2028+ idea, gated on Vision Pro 2 (or competitor) actually shipping reliable full-body tracking and a form factor athletes will wear under load. Don’t build it in 2026.

6. What a competent 2026 MVP for Hyrox actually looks like

Stack:

iOS native app, iPhone 14+ as minimum
On-device MediaPipe Pose (or Apple VNDetectHumanBodyPose3DRequest for iPhones with LiDAR)
Local rule engine for rep counting + joint-angle thresholds (knee valgus, depth, trunk lean) — borrow from the open-source fitness-trainer-pose-estimation patterns (GitHub)
10–30s clip is sent to Gemini 2.5 Flash after each set with a structured prompt: “This is a Hyrox wall ball. Score depth, full extension, ball-to-target. Reply in 2 sentences.”
Voice playback of the summary between sets
All clips kept locally; user can scroll back and see the worst rep

Movements covered at MVP launch: Wall Balls, Burpee Broad Jumps, Sandbag Lunges. Add Sled Push (form-only, distance from a watch), Farmer’s Carry posture, then SkiErg and Rowing as v2 hardware-integration features (read from C2/SkiErg’s BLE).

Feedback mode: B (between-set). Skip A and skip C. B is where the tech is, B is where users tolerate latency, B maps to how Hyrox training actually happens (interval-based).

Dev cost to MVP: One iOS engineer + one ML/CV-fluent person, 4–6 months. ~$80–120K all-in if contracted; less if founder-built.

Inference cost per user: At 3 sets/movement × 3 movements × 30s × Flash pricing ≈ $0.027 per session. A user training 4×/week costs ~$0.43/month in API fees. Even with $20/mo subscription, gross margin survives easily.

Realistic accuracy / satisfaction: Rep counting 95%+. Form flagging 70–80% true-positive on the three start movements with the camera placed correctly. User satisfaction depends almost entirely on (a) how forgiving the UX is about camera placement, and (b) whether the verbal feedback feels like a coach or a robot. Tonal’s failure mode is the latter — generic cues that don’t earn trust.

7. Honest risk assessment

Why a 2026 build might succeed where 2020–2024 didn’t

Multimodal LLM video understanding is genuinely new. No one had Gemini 2.5 Flash or GPT-4o video reasoning in 2022. The verbal feedback can now be coach-quality in a way that hand-coded cue libraries could never be.
Hyrox is a narrow, well-defined domain. Eight movements with known failure modes. Mirror tried to be all-fitness; Tonal tried to be all-strength. Narrow beats broad in vision-based fitness.
The community already records. Hyrox athletes already film their training. The MVP doesn’t have to convince anyone to point a phone at themselves — that behavior exists.

Why it might still fail

Camera placement remains the single largest UX failure mode. Onyx died on this. Every successful pose-based fitness product either solved this with proprietary hardware (Tonal, Tempo, Mirror’s box) or narrowed to a static framing (HomeCourt: phone on tripod facing the hoop). Indie founders can’t ship hardware.
Trust collapses on the first wrong call. A single false “your knee is caving” when it wasn’t ends the relationship. The 70–80% true-positive rate above is generous and may be lower in real gyms with bad lighting and partial occlusion from racks/sleds.
“Form coach” might not be the value users actually pay for. Look at HomeCourt: it survived by being a measurement tool, not a coaching tool. Hyrox athletes might pay $20/mo for accurate rep counting + race-pace prediction + leaderboards. Form correction may be a feature, not the product.

Brutal TL;DR

Highest-probability indie-founder Hyrox feature to ship in 6 months:

A measurement-first iOS app that uses on-device MediaPipe + Apple body pose to count reps and time the 8 stations accurately during a Hyrox-format session, with between-set qualitative feedback from Gemini 2.5 Flash on three movements (Wall Balls, Burpee Broad Jumps, Sandbag Lunges), and a voice summary at the end of each block.

Actual user experience it would deliver: “It accurately counts my wall balls and tells me my squat depth was inconsistent on reps 60–80. After my burpee broad jump set, it tells me my jumps shortened by 30% in the last 20m. It does not coach me mid-rep, it doesn’t catch every form error, and on the sled push and SkiErg it’s basically a fancy timer. But it’s the first app that knows what a Hyrox session looks like and gives me data I’d otherwise need a coach to capture.”

That product is buildable, defensible, and honest about what AI can do in May 2026. The “real-time AI personal trainer that corrects every rep” — the thing that killed Mirror, Onyx, and Vi — is still 2–4 years away, and is probably the wrong product anyway.