There's a lot of potential here, especially since you're already building custom firmware/controller logic and have deep experience with local AI. The torque sensor gives you a real-time force signal at ~100–200 Hz — that's a rich data stream most e-bike brands barely use beyond simple proportional assist. adventuro
Run a lightweight model (TinyML / TFLite on the controller MCU, or on a companion ESP32) that learns your pedaling signature over time: tamobykesport
- Builds a personal torque-cadence-speed profile across different terrains and conditions.
- Automatically adjusts the assist curve without you touching PAS levels — it just knows you want more on climbs and less on flats.
- Could run entirely offline, fits Omera's local-first philosophy. adventuro
Fuse torque sensor data with GPS + altimeter to anticipate what's coming: tamobykesport
- Pre-load motor torque 1–2 seconds before a hill hits, so assist feels seamless.
- Learn your commute routes and pre-cache optimal power curves per segment.
- Optimize battery spend: "You have 14 km left, 2 hills — here's the power budget."
This is the novel one. The torque signal contains subtle patterns that shift as you fatigue: hezzobike
- Pedal stroke asymmetry — left vs. right leg force imbalance increases when tired.
- Torque variability — consistent riders get jerky when fatigued.
- Cadence drift — natural RPM drops without the rider noticing.
- An on-device model could detect fatigue onset and silently increase assist to keep you safe, or nudge you to take a break.
Log torque + cadence + speed at high resolution and pipe it to a companion app: smartybikes
- Power output estimation (torque × cadence = watts) without needing a power meter.
- Stroke-by-stroke analysis like what cyclists pay €500+ for with Garmin/SRM.
- "Ghost mode" — ride against your own previous best on a route.
The torque signal is surprisingly biometric: hezzobike
- Everyone has a unique pedal stroke signature (force curve shape, L/R balance, cadence preference).
- Train a small classifier on your stroke pattern.
- If someone else pedals the bike → motor doesn't engage. No keys, no app, just pedal and it recognizes you.
| Feature | Complexity | Hardware needed |
|---|---|---|
| Adaptive assist curve | Medium | Torque + cadence + existing controller |
| Predictive terrain | Medium-high | + GPS module (ESP32 + u-blox) |
| Fatigue detection | Medium | Torque signal only, needs ML model |
| Power analytics | Low | Just logging + app |
| Rider authentication | Low-medium | Torque signal + small NN on MCU |
The power analytics + adaptive assist combo is probably the best bang-for-buck for v1 — useful immediately, differentiating vs. every other e-bike, and you can ship it with just the torque sensor you already have plus a small local model running on the controller. The rider-auth feature is a fun flex that fits Omera's anti-theft story too. tamobykesport
Want me to spec out the data pipeline for any of these (sensor → MCU → model → motor control)?
For your Omera e-bike controller running TinyML workloads (adaptive assist, fatigue detection, rider auth) alongside real-time motor control and BLE to the iPhone, the ESP32-S3 is the strongest all-around pick. Here's how it stacks up against the other realistic candidates.
ESP32-S3 — Best All-Rounder (Recommended)
The S3's headline feature is its vector processing extensions (PIE) on dual Xtensa LX7 cores at 240 MHz, which accelerate quantized neural network operations by roughly 40× compared to unaccelerated inference. Key specs:dronebotworkshop+1
512 KB SRAM + optional 2–8 MB PSRAM
Built-in Wi-Fi 4 + BLE 5.0 (no separate radio chip needed for iPhone comms)
Rich ADC/SPI/UART for torque sensor input and motor PWM output
Proven TFLite Micro support: ~50–60 ms inference, ~240 KB model size, ~350 KB RAM usage for a quantized CNN[dev]
Costs ~€3–5 per chip, massive community and tooling
This comfortably handles your small classifier models (rider auth, fatigue detection) at 100–200 Hz sensor rates while leaving headroom for motor control on the second core.[zediot][ppl-ai-file-upload.s3.amazonaws]
Nordic nRF5340 — Best for Ultra-Low Power + BLE Quality
If you prioritize best-in-class BLE and ultra-low power consumption over raw ML throughput, the nRF5340 is worth considering:[jonpeddie]
Dual Arm Cortex-M33 (128 MHz app core + 64 MHz network core)
Excellent BLE 5.0 with Thread/Zigbee multi-protocol
Edge Impulse partnership for rapid TinyML deployment — gesture models run in 1.7 ms using ~4 KB[hackster][youtube]
Lower ML ceiling than the S3, but sufficient for small sensor classifiers
The trade-off is less community momentum for motor-control applications and a more constrained toolchain compared to ESP-IDF.[edgeimpulse]
ESP32-P4 — Future Upgrade Path
The P4 is Espressif's 2026 powerhouse: dual RISC-V at 400 MHz, 32 MB PSRAM, AI instruction extensions, and hardware H.264 encoding. However:linkedin+1
It requires a separate ESP32-C6 co-processor for Wi-Fi/BLE — adds board complexity
Designed for vision/multimedia, which is overkill for torque sensor ML
ESP-IDF support is mature but Arduino support is still limited[dronebotworkshop]
This makes sense if Omera v2 adds a camera (e.g., dash-cam or gesture control), but it's overengineered for v1.
STM32N6 — Overkill
The STM32N6 packs a dedicated Neural-ART NPU delivering 600 GOPS at 800 MHz Cortex-M55. It's designed for on-chip computer vision and audio — far beyond what torque-sensor classifiers need, and significantly more expensive.[embedded]
Recommendation for Omera v1
Go with the ESP32-S3 (N8R2 or N16R8 variant) for Omera v1. It gives you TinyML inference, BLE to iPhone, motor control GPIOs, and GPS module support (via UART to a u-blox module) — all on a single chip at a few euros. The N16R8 (16 MB flash, 8 MB PSRAM) gives you generous headroom for model storage and data buffering. You already know ESP-IDF from your Arduino/embedded work, so the ramp-up time is minimal.github+1
ESP32-S3 — Best All-Rounder (Recommended)
The S3's headline feature is its vector processing extensions (PIE) on dual Xtensa LX7 cores at 240 MHz, which accelerate quantized neural network operations by roughly 40× compared to unaccelerated inference. Key specs: dronebotworkshop
This comfortably handles your small classifier models (rider auth, fatigue detection) at 100–200 Hz sensor rates while leaving headroom for motor control on the second core. [zediot](https://zediot.com/blog/esp32-chip-series-comparison/)
Nordic nRF5340 — Best for Ultra-Low Power + BLE Quality
If you prioritize best-in-class BLE and ultra-low power consumption over raw ML throughput, the nRF5340 is worth considering: [jonpeddie](https://www.jonpeddie.com/news/nordic-semiconductor-expands-its-ai-footprint/)
The trade-off is less community momentum for motor-control applications and a more constrained toolchain compared to ESP-IDF. edgeimpulse
ESP32-P4 — Future Upgrade Path
The P4 is Espressif's 2026 powerhouse: dual RISC-V at 400 MHz, 32 MB PSRAM, AI instruction extensions, and hardware H.264 encoding. However: linkedin
This makes sense if Omera v2 adds a camera (e.g., dash-cam or gesture control), but it's overengineered for v1.
STM32N6 — Overkill
The STM32N6 packs a dedicated Neural-ART NPU delivering 600 GOPS at 800 MHz Cortex-M55. It's designed for on-chip computer vision and audio — far beyond what torque-sensor classifiers need, and significantly more expensive. embedded
Recommendation for Omera v1
Go with the ESP32-S3 (N8R2 or N16R8 variant) for Omera v1. It gives you TinyML inference, BLE to iPhone, motor control GPIOs, and GPS module support (via UART to a u-blox module) — all on a single chip at a few euros. The N16R8 (16 MB flash, 8 MB PSRAM) gives you generous headroom for model storage and data buffering. You already know ESP-IDF from your Arduino/embedded work, so the ramp-up time is minimal. github