These methods are designed for training a vehicle follower, which we plan on expanding to platooning. However, we might expand this to train a single racer in the future.
Both methods require a parameter file is loaded before running. We'll explain these more below, but they contain all of hyperparameters for the learning algorithm. Modifying these can create different results which might be better than our performance. We use nominal values discussed in papers opting for consistent results instead of peak performance.
Deep Deterministic Policy Gradient (DDPG) is explained in Lillicrap et. al, 2015. The method uses model-free, off-policy learning to determine an optimal deterministic policy for the agent to follow. This means, during training, the agent executes random actions not determined using the learned policy. This string of randomly executed actions is stored in a replay buffer, which is sampled from after each step to learn an estimation of the state-action value (a.k.a. Q-value) function. This Q-function is used to train the policy to take actions that result in the largest Q-value. The folder containing this method, DDPG, is made of the following parts:
- config_ddpg.yaml: YAML file with all of the configuration information including the hyperparameters used for training.
- ddpg.py: The main file for running this method. Every step of the DDPG algorithm is implemented in this file.
- nn_ddpg.py: The neural network architecture described in Lillicrap et. al, 2015 is implemented here using PyTorch. Forward passes and initialization are handled in this class as well.
- noise_OU.py: Ornstein-Uhlenbeck process noise is implemented in this file. This noise is applied to create the random exploration actions. This is the same noise used in Lillicrap et. al, 2015.
- replay_buffer.py:
To run this method, do the normal launch of the racing environment with 2 cars, then in a separate terminal:
$ rosparam load config_ddpg.yaml
$ rosrun rl ddpg.pyProximal Policy Optimization (PPO) is explained in Schulman et. al, 2017 as a simplified improvement to Trust Region Policy Optimization (TRPO). The method uses model-free, on-policy learning to determine an optimal stochastic policy for the agent to follow. This means, during training, the agent executes actions randomly chosen from the output policy distribution. The policy is followed over the course of a horizon. After the horizon is completed, the Advantages are computed and used to determine the effectiveness of the policy. The Advantage values are used along with the probability of the action being taken to compute a loss function, which is then clipped to prevent large changes in the policy. The clipped loss is used to update the policy and improve future Advantage estimation. The folder containing this method, PPO, is made of the following scripts:
- class_nn.py: The neural network architecture described in Lillicrap et. al, 2015 is implemented here using PyTorch. Forward passes and initialization are handled in this class as well. We are currently working on changing the architecture to match that used in Schulman et. al, 2017.
- ppo.py: The main file for running this method. Every step of the PPO algorithm is implemented in this file.
- ppo_config.py: YAML file with all of the configuration information including the hyperparameters used for training.
To run this method, do the normal launch of the racing environment with 2 cars, then in a separate terminal:
$ rosparam load ppo_config.yaml
$ rosrun rl ppo.pyIf the car crashes or you want to start the experiment again. Simply run:
$ rosrun race reset_world.pyto restart the experiment.
To run a node to tele-operate the car via the keyboard run the following in a new terminal:
$ rosrun race keyboard_gen.py racecar'racecar' can be replaced with 'racecar1' 'racecar2' if there are multiple cars.
Additionally if using the f1_tenth_devel.launch file, simply type the following:
$ roslaunch race f1_tenth_devel.launch enable_keyboard:=true