Aster

An AI exercise-review prototype for people training alone, combining observations from workout videos with ORBIT's spoken feedback and a first-person space adventure.

Comment

Track: AI + Healthcare

Aster explores how AI can make exercise information easier to understand and use for people who are training alone. The aim is accessible support for physical wellbeing through feedback on a person's own workout footage.

Inspiration

I was alternating pull-up sets with someone when I noticed how deliberately he moved. I asked him to watch my form. We started talking, I could ask questions, and I got to learn from someone who was paying attention.

What stayed with me was the feeling of having someone there. That is the experience I want ORBIT, Aster's companion, to grow toward: a gym buddy who helps you understand your movement and a useful next step.

My target users are people exercising independently in places such as gyms and parks, particularly those who do not regularly have someone to ask about what they are doing.

The experience I am building

Aster places exercise review inside a first-person world on a distant moon. You explore the landscape, discover an exercise park, and meet ORBIT, a robot guide.

The setting comes from my fascination with space. I love the curiosity and perspective I associate with films such as Interstellar and The Martian, and the immersion I have felt playing games such as God of War. Learning about fitness also involves discovery: trying, noticing, adjusting, and finding your own direction. I wanted that idea to run through the whole experience.

Current functionality

The prototype supports recorded videos of pull-ups, push-ups, and squats. Its review interface can show tracked body landmarks over a recording, surface movement observations, and present spoken feedback through ORBIT alongside written details.

The current output mainly describes visible movement. The full journey is still inconsistent, and improving the feedback from observations into specific, dependable guidance is an active development priority. Saved workout history and sustained progress comparisons are future work.

How the AI works

A pretrained MediaPipe pose model analyses video locally in the browser. The application derives measurements and uses rules to assemble movement observations. Deepgram converts the resulting feedback text into ORBIT's speech. It provides the voice; the pose model and application logic provide the movement analysis.

The exercise video stays in the browser, while feedback text is sent to the speech service. TypeScript, React, Vite, and Babylon.js support the interface and three-dimensional world.

I worked solo with Codex assisting extensively with coding and implementation. The concept, creative direction, product decisions, testing, and iteration were mine.

What I learned and why it matters

Tracking a joint is only a starting point. A useful companion must help someone understand why an observation matters. The project includes an earlier 600-clip evaluation of the processing pipeline; that work does not establish form-correction accuracy.

The opportunity I want to pursue is clearer, more approachable exercise information for people practising on their own. My next steps are a reliable complete review, evaluation of the guidance with qualified exercise professionals, and practice history that makes progress easier to understand.

Build history

Aster's concept and initial planning began on 2 October 2026, before the ForgeHacks build window. The supplied Git history records implementation commits from 7-9 October covering the playable exercise environment, local movement analysis, pose evidence, evaluation refinements, and spoken workout reviews.

Built With

Share this project:

Updates

Submission history