E.L.I.S.E. - Enhanced Lacrosse Intelligence for Shot Evaluation

I chose to do this project because I noticed a frustrating pattern in my own lacrosse practices. I would repeat the same mistakes over and over, and by the time a coach corrected me, the mistakes had already become a bad habit that was hard to break. After talking to my teammates, I realized we all shared the same issue that in a two hour practice, we might only get one or two small pieces of feedback.

It is incredibly difficult for anyone to pick out a specific technical flaw in a shot that happens in the blink of an eye. My project fills that gap by catching the tiny details that a human coach might miss in the middle of a fast-paced practice.

Noticing these problems inspired me to create an AI coach that could communicate with players in real-time. My goal was to make this technology as affordable as possible and design it so that it doesn't require another person to operate. By providing instant data, I want to create an environment where athletes can thrive, correct themselves, and see progress every single day.

Purpose. Build a model that predicts the speed of a lacrosse shot from accelerometer and gyroscope data captured by a sensor mounted to a lacrosse stick.

Hypothesis. An accelerometer and gyroscope mounted on a lacrosse stick can predict shot speed within ±5 km/h of a radar gun.

Sensor Selection

The peak of a lacrosse show lasts only 10-30ms. I wanted an IMU that samples at 1000 hz to ensure I captured the peak.

Has a greater than ±16g force range, because hard shots can easily exceed 16g

Accelerometer and gyroscope (to find best predictor)

I picked the TDK ICM-45686 because it satisfied my initial requirements.

Also considered:

iPhone IMU : ±16 g, too big, can't attach to a stick.

MPU-6050: only ±16 g

Circuit Design Components

$32 total:

LilyGo T-Display S3, $18

TDK ICM-45686 breakout, $12

300 mAh LiPo, $2

Breadboard prototype: On my original prototype, loose wires added noise to the sensor data and sometimes disconnected. I soldered everything to a perfboard to fix the noise and make it smaller.

3D-Printed Stick Mount

Mounting a sensor on a fast swinging stick is tricky.

Four iterations:

Flat bottom box: wobbled on the octagonal shaft.

Curved bottom: shifted during shots.

Triangular grip: less wobble, but slid down the shaft.

Final design: rectangular case with screen cutout, board/battery slots, corner triangles for the octagonal shaft, zip-tie slots

iOS App

connects to the sensor using WiFi to retrieve shot data

logs participant name, radar gun speed, notes

graphs the swing acceleration and angular velocity

Stick Sensor

reads all 6 IMU channels at 1000 Hz

each shot is stored on the MCU

streams it to the iPad over WiFi after the shot

Both the iPad app (Swift) and Sensor software (C++) were written with assistance from Claude Code.

Data Gathering Procedure

Athlete takes three warm-up shots.

Stands 7.3 m from the goal line.

The instructor aims the radar gun and starts recording.

The athletes takes the hardest shot they can while hitting the net.

Five successful shots per athlete, 15 attempts max.

Results

I collected 122 shots over 28 athletes. The fastest shot was 110 km/h and the average was 52 km/h. 30 of those shots (24.6%) exceeded the accelerometer's ±32g limit and got clipped. They were always the fastest shots, averaging 77 km/h, compared to 43 km/h for the non-clipped shots.

On a really hard swing, the accelerometer returns the maximum value of 32g no matter what the true acceleration was. This means a 110 km/h shot and a 60 km/h shot would look like they have the same acceleration reading. To address this, I ran my analysis two ways: once on the the full 122-shot set (including clipped shots), and once on the 92-shot set with no clipping.

Analysis

Each shot generated over 5,000 readings (five or more seconds) across 3 axes of acceleration and 3 axes of rotation. I extracted the peaks, and combined the sensor data across all axes to also find the peak magnitude for acceleration and gyroscope.

That gave me 8 different speed predictors to compare. For each predictor, I found the line of best fit using least squares regression and tested it with leave-one-out cross-validation on the 92 clean shots. For every shot, I refit the line on the other 91 and used it to predict the held-out one. The model never sees the test shot while training, so it can't just memorize the data. The MAE is the average absolute error across all 92 held-out predictions. Out of all the prediction formulas peak acceleration magnitude was the winner with the lowest MAE of 4.76 km/h.

Impact of clipping

The main problem with accuracy is the accelerometer's range. Accelerometer based predictions get much better once you remove the clipped shots.

All 122 shots: 8.43km/h MAE

92 clean shots: 4.76km/h MAE

Peak acceleration magnitude's prediction error is reduced by 43% without clipped shots. A higher range accelerometer (±64g or more) could catch all the detail of those clipped shots and allow for accurate prediction of much harder shots.

Conclusion

The results support the hypothesis. An IMU mounted on a lacrosse stick combined with software can predict a lacrosse show speed with a MAE of 5km/h

Accomplished

Designed and built a $32 sensor that is able to capture lacrosse shot dynamics at 1000 Hz

Designed and 3D printed a case that protects the sensors that clamps onto a lacrosse stick and that withstands 32+ g forces during shots

Demonstrated the sensor predicts shot speed within 4.76 km/h (non-clipped) across 28 players and 122 shots

Developed an iPad app for real-time data collection and shot analysis

Coach E.L.I.S.E.

Real-time spot speed prediction

AI coaching after every shot

Session check-ins every 5 shots

Five selectable coach personalities

The personalities are:

Coach Elise (encouraging)

Sergeant Shot (sarcastic)

Pro Coach (technical)

Hype Machine (excited)

Coach Zen (calm)

Higher-range accelerometer: I have a new prototype ready using an accelerometer (ADXL372) capable of 200g.

Smaller case: The current case is a bit too big and throws off the stick's balance. My new prototype uses a smaller MCU (ESP32-C6) and will be half the size.

Other sports: Investigate whether the same approach of using a sensor to analyze swings works for baseball, tennis, golf, and hockey.

Combined IMUs: Connect two IMUs inside the case, one to be able to handle high G-forces and the other to capture a more detailed graph of the shot.

I would like to thank all of the wonderful athletes who took part in my experiment. Halifax Independent School was also a huge contributor by providing their gym space so I could complete all of my trials.

LILYGO. (n.d.). LILYGO T-Display-S3 ESP32-S3 development board with 1.9 inch LCD, 16MB Flash, and 8MB PSRAM [Product page]. Amazon. https://www.amazon.ca/LILYGO-ESP32-S3-Development-

Wireless-Module/dp/B0BLMNQRTLLILYGO. (n.d.).

LILYGO T-Display-RP2040 Raspberry Pi development board with 1.14 inch LCD [Product page]. Amazon. https://www.amazon.ca/LILYGO%C2%AE-T-Display-RP2040-Raspberry-Development/dp/B09J112YR7

ProjectBoard. (n.d.). Modeling human movement: Scalable biomechanical analysis with low-cost wearables and machine learning. Youth Science Canada. https://partner.projectboard.world/ysc/project/modeling-human-movement-scalable-biomechanical-analysis-with-low-cost-wearables-and-machine-learning

Tinkercad. (n.d.). Amazing Gogo-Amberis [Circuit design]. Autodesk. https://www.tinkercad.com/things/dyIhqoInxJm-amazing-gogo-amberis

Volos Projects. (2021, September 24). LilyGo T-Display RP2040: Review and tutorial [Video]. YouTube. https://www.youtube.com/watch?v=WGp07cI4jZg

Camomilla, V., Bergamini, E., Fantozzi, S., & Vannozzi, G. (2018). Trends supporting the in-field use of wearable inertial sensors for sport performance evaluation: A systematic review. Sensors, 18(3), 873. https://doi.org/10.3390/s18030873

Antropic

King, K., Yoon, S. W., Perkins, N. C., & Najafi, K. (2008). Wireless MEMS inertial sensor system for golf swing dynamics. Sensors and Actuators A: Physical, 141(2), 619–630. https://doi.org/10.1016/j.sna.2007.10.028

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay, É. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12(85), 2825–2830. https://www.jmlr.org/papers/v12/pedregosa11a.html

Random Nerd Tutorials. (n.d.). ESP32 I2C communication (Arduino IDE). https://randomnerdtutorials.com/esp32-i2c-communication-arduino-ide/

Random Nerd Tutorials. (n.d.). ESP32 with MPU-6050 accelerometer, gyroscope and temperature sensor (Arduino IDE). https://randomnerdtutorials.com/esp32-mpu-6050-accelerometer-gyroscope-arduino/

TDK InvenSense. (2023). ICM-45686 datasheet (Rev. 1.2). https://invensense.tdk.com/wp-content/uploads/2023/06/DS-000494-ICM-45686-v1.2.pdf

tdk-invn-oss. (n.d.). motion.arduino.ICM45686 [Source code]. GitHub. https://github.com/tdk-invn-oss/motion.arduino.ICM45686

Coding Assistant

Anthropic. (2025). Claude Code [Computer software]. https://claude.com/claude-code