lexfridman/deeptraffic

Competitors design neural network architectures and tune hyperparameters (layer count, learning rate, discount factor) via an in-browser IDE to train DQN agents that navigate a 7-lane highway simulation using occupancy grid inputs. The platform provides real-time visualization of network activations and reward signals, enabling crowdsourced exploration of motion planning hyperparameter spaces for multi-agent autonomous driving scenarios where each vehicle independently executes the same learned policy.

1,793 stars. No commits in the last 6 months.