MaLPi Task List

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.

1. ✓ Pre-train VAE on images from all but one track
   • Add auxiliary outputs: CTE, track #, steering, throttle
   • Collect more training data from all tracks
2. ✓ Train separate policies on each track with the VAE as an image embedder
   • Testing script needs to output a single value with standard deviation
3. Implement Pre-training paper
4. Switch to RNN policy
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. Switch to more formal multi-task and/or meta-learning and/or lifelong learning
8. Add some form of Aggregate memory that includes all previous experience
9. Add an IMU and learn to detect crashes/bumps

Finished

1. ✓ VAE based DonkeyCar pilot
2. Switch from fastai to Lightning
3. Train a single policy on multiple tracks/tasks

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.