Neural Network Imitator

[Florian-Bolli]

ToDo:

• [x] Make NNI planner automatically adjust to required neural network inputs [@Florian-Bolli ]
• [ ] Train with control and state noise on different maps
• [x] Finetune pretrained (with noise) network with data (without noise) [@nigalsan1]
• [ ] Finetune pretrained (with noise) network with data from physical map (ETF2) [@nigalsan1]
• [x] Finetune pretrained (with noise) network with pure pursuit [@nigalsan1]
• [ ] Train Neural Network entirely without noise [@Florian-Bolli ]
• [ ] Train NN with pure pursuit for low speed and MPPI for higher speed
• [ ] Analyze data from physical car collected on ETF2
• [ ] Train networt with noise MPPI on RCA1 and finetune with PP [@nigalsan1]
• [ ] Implement a Recording for laptime for pyhiscal racing

Longterm:

• [ ] Think of way how to avoid obstacles with non reliable lidar data [@nigalsan1 , @Florian-Bolli]
• [ ] Analyze lidar data from physical car on diferent maps with and without obstacles
• [ ] NNI planner following waypoints generated by MPC
• [ ] obstacles detection with a second network which is handling some LiDAR points and inputing the rays, which detected something into the wp network

Other ideas:

• [ ] train with varying friction coefficient to make if tuneable to differen floors

[nigalsan1]

I have tried 2 trainings data generation without noise and trained all the models and the had weird behavior in sim It seemd like they forgot some actions during training and try to overreact in the curves