ML
Training
- Finish training pipeline
Deployment
- Local deployment as a ROS node/package with checkpoint onboard
- Cloud deployment
- Remote ROS node running on Paperspace with ROSmaster URL
- Endpoint – Azure, AWS, Cerebrium, RunPod
- Message schema and passing mechanism to mapping system
- Multiple deployment methods?
- Basically I want deployment to be robust AND easy to use but local and cloud deployment methods each have flaws
- Local: Different hardware for inference, CUDA…
- Cloud: Access to internet? (Kodama, IronTree, Boat)
Bring-up
- Entire ML pipeline should have a clearly defined and smooth bring-up process, eventually become a part of Polymath’s general bring-up
- Maybe separate the parts we need from the big MMDet3D repo
- Is it easier for me to create in-house tools for collection/labeling?
- Data collection → Labeling → Training → Deployment
Other
Neglected (?) Projects
- Automated EKF Tuning: Thought it might be interesting to pick up some work in this area, Alex doesn’t seem to have too much time to play around with it but I’m also not too sure what needs to be done
- Good opportunity for me to learn more about EKF + MLE
- Baymax Robot Health Monitor: Heard this mentioned a couple times but don’t think there’s much work being done on it, maybe I can help
Robots/General Engineering
- Be more involved in core autonomy development for clients, robot bring-up
- Will is gone so maybe pick up some slack on his projects
Work hard + have fun!
- ML Deployment
- Fuse bring-up
- Covariance injector for fuse
- Baymax
- Compound Eye
Baymax: