Skip to main content

Hi Everyone! Great reading all of the other excellent ideas for the pioneer 10 project. A big thank you to Axelera for creating this challenge.

My project is for an AI powered gym buddy, which i’ve decided Brobot is a sensible name…

When doing gym workouts, its common to aim for 10 repetitions of a movement. If you want to progress the last reps should be a struggle. If it’s not you should push some more reps, or next time add more resistance/weight. That is how you get better over time. It’s called progressive overload.

Just last week, I pushed to much and pulled something in my neck while doing overhead press (pressing two dumbells above my head). It would have been clear in a mirror by body was twisting and straining to push the reps - this is bad form and is a too common cause of injury.

A trainer would have told me to stop pushing because the form was bad.

I think it’s totally feasible to have a camera watching the movements, checking form with pose models, then giving the user some feedback on when it’s ok to push a little more or to stop as soon as the form is looking suspicious.

From an IO point of view i’m keeping it simple, there will be audio and visual alerts via an Arduino, and a HDMI display for user feedback.

I think I should be able to get plenty of “good” examples of reps online, the first step will be data collection and validating what can be done with ready pose detection models. My core focus will be learning the Voyager SDK.

Awesome idea for a project. And one I look forward to following. Quick question that sprung to mind after reading the project introduction. Could a model be trained on existing fitness video content for form?


Thanks for your question! Yes, I think the bulk of the training could be done with existing content. In fact i’m really hoping it is, otherwise I have a lot of gym time coming up...


Really great idea, and I love the new name ​@llrds😄💪 The truth is, mirros in a gym aren’t just for posing - they have a practical value! So this is a great safety application, not just for improvement of form. Really looking forward to it dude!


I have made a PoC with OpenCV so I can try out the models, so far it’s been good results from even the nano pose model, I look forward to to getting the proper hardware to try the larger models! The next challenge I am working on now is classification, while I don’t think I need to train the existing model I still will need to collect data on workouts.

I have written up what I have done so far on my LinkedIn.

https://www.linkedin.com/pulse/first-steps-ai-model-selection-process-liam-layton-l6hfe


Reply