In this project, we aim to take our SpiderBot and make it walk using deep reinforcement learning. The code is written entirely in Python 3.7.7 and the following Python libraries are required for our code to work. The code can be found here
pybullet==3.0.6 numpy==1.18.5 matplotlib==3.3.2 tensorflow_probability==0.11.1 seaborn==0.11.0 pandas==1.1.4 tensorflow==2.3.1
Other than this, no additional software is needed for the code to work. The PyBullet Physics Engine is used for simulation using an OpenGL GUI. In this code, we have the following : -
- A requirement.txt for required python libraries
- SolidWorks CADs of the SpiderBot
- SpiderBot URDFs for the SpiderBot
- Folders for Training Logs & Plots
- Two saved models of the SpiderBot Agent
- Source Code for the Deep RL Implementation
- Training Code to train the SpiderBot with Deep RL
- Validation Code to test trained models
- Postprocessing Code to generate plots of training
The code supports the following 5 algorithms (with their characteristics defined):
|Algorithm||Agent (Actor)||Policy||Learning Network||Actions per Time-Step||Action Space||State Space|
We will now walk through the folders and files.
This folder contains all the part and assembly files for the SpiderBot. There are options for 3-legged, 4-legged, 6-legged & 8-legged SpiderBots.
This folder contains all URDF files and associated STL files for the SpiderBot. There are options for 3-legged, 4-legged, 6-legged & 8-legged SpiderBots.
Training_Logs & Training_Plots
Folders to store csv file of training data and PDF plots of training.
Contains two saved models using DDPG. The FullyTrained Model (375 episodes) is able to walk well and up to 9 metres in the forward direction. The PartiallyTrained Model (50 episodes) can move forward slightly but only to a certain extent.
This file has the p_gym class. This uses pybullet and loads the plane environment (no obstacles) and the SpiderBot into the physics engine. The code allows an agent to retrieve state observations for a leg or whole SpiderBot and set a target velocity for joints in the SpiderBot. Finally, the code uses information from the physics engine to determine rewards for a time step.
This file has the classes for the fully-connected neural networks used. The Tensorflow 2 API is used to develop the neural networks. Depending on the algorithm and number of SpiderBot legs, the neural networks are customised for them. There is all a call method to do a forward propagation through the neural network.
This file is a long one, which has all the operations of the agent for all 5 algorithms. It initialises the neural networks based on the algorithm in the constructor. The class also has the functionality to update the target networks for DDPG & MAD3QN. Additionally, it has a long list of methods to apply gradients for each one of the algorithms. In these methods, the TensorFlow 2 computational graph and gradient tapes are used to help in backpropagating the loss function. Finally the class also has the functionality to save all models and load all models.
This file contains the replay_buffer class that handles experience replay storage and operations like logging and sampling with a batch size.
This file contains the walk function that is actually responsible for handling all training operations. This is where all the classes interact with each other. The episodes are looped through and the SpiderBot is trained. The training-related data is logged and saved as a csv into the Training_Logs folder while the best models are saved to the Saved_Models folder during training.
This file handles the plotting post-processing operations that takes the CSV file from the Training_Logs folder and saves the plot into the Training_Plots folder.
This file allows the user to set up the training session. In this file, the user can set 3 levels of configuration for training. The general config section has options for choosing algorithms, number of legs, target location, episodes etc. The Hyperparameters config section handles all hyperparameters of the entire training process. The reward structure config provides options for all the scalar rewards. The user must set all of these configs and run the file to train the SpiderBot. TIP: not using a GUI is faster for training, especially if you use a CUDA-enabled NVIDIA GPU.
This file allows the user to validate and test a trained model, specially made for the Professors and TAs of SpiderBot to visualise our fully trained model.
How to train a model?
Unzip the SpiderBot_URDFS.zip file into the same directory. Open up
SpiderBot_Train_Model.py for editing. The most important parameter is
training_name that you must define. This is unique to a particular training session and all saved models, logs and plots are based on this
training_name. After that set up your General Config:
#~~~~~~~~~~~~~~~~~~~~~~~~~~~ GENERAL CONFIG ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~# training_name = "insert_training_name_here" model = "DDPG" num_of_legs = 8 episodes = 375 target_location = 3 use_GUI = True do_post_process = True save_best_model = True save_data = True
Following that, set up the configurations for the hyperparameters:
#~~~~~~~~~~~~~~~~~~~~~~~~~~~ HYPERPARAMETER CONFIG ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~# time_step_size = 120./240 upper_angle = 60 lower_angle = -60 lr_actor = 0.00005 lr_critic = 0.0001 discount_rate = 0.9 update_target = None tau = 0.005 max_mem_size = 1000000 batch_size = 512 max_action = 10 min_action = -10 noise = 1 epsilon = 1 epsilon_decay = 0.0001 epsilon_min = 0.01
Finally, set up the configuration for the reward structure:
#~~~~~~~~~~~~~~~~~~~~~~~~~~~ REWARD STRUCTURE CONFIG ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~# forward_motion_reward = 500 forward_distance_reward = 250 sideways_velocity_punishment = 500 sideways_distance_penalty = 250 time_step_penalty = 1 flipped_penalty = 500 goal_reward = 500 out_of_range_penalty = 500
Then run the python code
> python SpiderBot_Train_Model.py
How to Validate/Test our Models?
To test the fully trained model, just run
> python SpiderBot_Validation.py
If you wish to run the other saved model, the partially trained one, you can open up
SpiderBot_Validation.py and edit the
DDPG_PartiallyTrained in the config section as shown:
#~~~~~~~~~~~~ VALIDATION CONFIG SETUP ~~~~~~~~~~~~# training_name = "DDPG_PartiallyTrained" model = "DDPG" target_location = 8 episodes = 100000000000 # A large number is set to put the simulation on loop