Machine Learning on a Raspberry Pi (also a subtle reference to the Stargate shows)

A robotics project for playing around with hardware and learning to use Machine Learning techniques.

• An early autonomous lap with MaLPi v3

Task

This is my very tentative list of features to add to MaLPi. Although they are listed in numbered order, only the first one is ‘real’. If/when I finish the first task I’ll re-asses the list, possibly moving things around or dropping some or adding new ones. A checkmark means I’ve made some progress on that entry.

1. ✓ Switch to RNN policy
2. Start training with RL in adition to imitation learning
3. ✓ Output a single value with standard deviation from the testing script
4. VAE improvements
   • Add auxiliary outputs: CTE, track #, steering, throttle
   • Collect more training data from all tracks
   • Implement Pre-training paper
5. Add a small DNC as a working memory
6. Add more tasks
   • Cone/Human/dog/etc detector (bounding box as output)
   • Drive to goal with goal given as an image
   • Sketches as inputs (goals) and/or outputs
   • NLP description of a scene or task trajectory as output
7. Add more input and output modalities
   • Add an IMU and learn to detect crashes/bumps
8. Switch to more formal multi-task and/or meta-learning and/or lifelong learning
9. Add some form of Aggregate memory that includes all previous experience

Finished

1. VAE based DonkeyCar pilot
2. Train separate policies on each track with the VAE as an image embedder
3. Switch from fastai to Lightning
4. Train a single policy on multiple tracks

Model Predictive Control is a control method based on planning n-steps into the future to find the best path, as currently determined by the model. Take the first step. Then re-plan, take the new first step. Repeat until done.

Software

My GitHub repo
My DonkeyCar fork

Current Hardware

• Raspberry Pi 3 Model B $39.95
• Raspberry Pi Camera Board v2 - 8 Megapixels $29.95
  • Unlike a standard DonkeyCar I don’t use a wide-angle version.
• Raspberry Pi Camera Case for V2 8 Megapixel $8.49
• Unknown make/model battery $40?
  • Bought from the California Browncoats booth at Comic-Con.
  • Unfortunately, it does not support pass-through charging.
  • Only used to power the Raspberry Pi.
• Servo driver
• Exceed Racing Desert Monster 1/16 Scale Truck ~$90
  • This particular model may not always be available. Check the DonkeyCar website for up-to-date build instructions.
• Relatively standard DonkeyCar base plate and arch
  • Printed on an Ender 3.

Estimated total hardware costs: $210-$230

Previous Hardware Versions

• MaLPi v2
• MaLPi v1 was largely built using lego parts and spare motors I had lying around. It never got far enough along to actually drive.

Misc

Link to my Mastodon profile. (Should go in the side-bar)