Super Donkey Cars
summary
An open source Donkey car is rebuilt to train on an AWS GPU instance and then run real-time inference on a Raspberry Pi; hardware costs ~€220, data collection requires ~10–20 laps per track, default images are 160×120 (they try 320×240), and training uses TensorFlow that must be recompiled for ARM. The build fakes throttle, runs steering directly from the network, and fails to generalize when a table becomes a spurious cue; a PlayStation joystick produces abrupt labels that prevent convergence.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:04What is Donkey Car
- 2:50Assembling hardware
- 5:15Software setup
- 8:28Training the car
- 11:26First drive attempt
- 13:18Improving the model
- 15:49Debugging input and visualization
- 17:17Donkey Car accessibility
- 18:23Q&A: Input differences
- 20:07Q&A: Control output
- 21:31Q&A: RC car compatibility
- 22:56Q&A: Simulated training
- 25:06Q&A: Overfitting issues
- 27:14Q&A: Beyond track driving
- 28:38Q&A: Reinforcement learning
- 29:20Q&A: European community
- 30:46Q&A: Competitions and plans
- 31:28Q&A: Joystick issues
- 32:35Closing
related talks
35mDingoCar: Perception module Small scale, self-driving car
01h 25mHow to Build Your Own Self Driving Toy Car
37mAWS DeepRacer: Train, Evaluate, and Tune
37mAccelerate your team’s machine learning journey with DeepRacer
27mGet rolling with machine learning on AWS DeepRacer
29mGet rolling with ML on AWS DeepRacer
37mAWS DeepRacer:train, Evaluate, and Tune Your Reinforcement Learning Model
20mAccelerating AI pipelines w/ NVIDIA on AWS: An autonomous driving case