GatorPilot Preliminary Demonstration
Scene Text Detection
Bayesian Deep Reinforcement Learning for Self-Driving Safety
- DDPG Video : This video demonstrates the enhanced Deep Deterministic Policy Gradient algorithm (reference). A DDPG agent controls the steering angle of the car within TORCS car driving simulator. This DDPG agent outputs a continuous smooth steering angle decision, i.e., a steering angle that can vary between complete left and complete right, as well as acceleration and braking signals. With extensive training on a variety of track scenarios, we can see that the car can drive at near human-level performance.
- DQN Video : These videos demonstrate the Deep Q Network algorithm (paper) pioneered by Mnih et. al. We first tested the DQN agent in an Atari game “Enduro” where it performs well. We further test the DQN agent inside a 3D racing simulator. The DQN agent controls the steering angle of the car within TORCS car driving simulator. The DQN agent outputs a discrete coarse steering angle decision i.e. complete left, complete right or straight. It is seen the performance is very sub-par. The agent is barely able to drive the car a few meters after preliminary training.