Skip to content

a-mufasa/AWS-DeepRacer

 
 

Repository files navigation

Open in Visual Studio Code

AWS DeepRacer Final Project

National Action Council for Minorities in Engineering (NACME) Google Applied Machine Learning Intensive (AMLI) at the University of Arkansas

Developed by:

Project Description

According to the "Capstone Projects" pdf, which introduced us to the possible final projects, the goal for the AWS DeepRacer Project is to "train a 1/18th scale car to race autonomously using Reinforcement Learning."

Intro to Reinforcement Learning

Reinforcement Learning differs from the other 2 basic Machine Learning paradigms (Supervised & Unsupervised). A specific difference we can point to between Reinforcement Learning and Supervised Learning is the unnecessary input/output labelings as RL algorithms typically use dynamic programming techniques with the goal of automatically improving through reward maximization.

image

Dataset

The dataset our autonomous vehicle will use is the live sensor data it extracts which is used to train our model to determine the correct (highest reward) action via our reward function. This automatic trial-and-error process will give us a model that can respond to similar environments as our training with the same actions deemed correct by the algorithm.

Usage instructions

  1. Make sure you have access to the AWS Console for training & evaluating the model
  2. Follow the AWS DeepRacer instruction on how to 'Create a Model' and use the default hyperparameters on the 're:Invent 2018' track.
  3. Copy the code in final_reward.py
  4. Paste the reward function code into AWS and train & evaluate the model

Final Model: Training Configuration

re:Invent 2018 track

Training happens through an iterative process of simulation to gather experience, followed by training on the experience to update your model, followed by simulation using the new model to get new experience, followed by training on the new experience to update your model and so forth. Initially your model does not have any knowledge of which actions will lead to good outcomes. It will choose actions at random as it explores the environment. Over time it will learn which actions are better and start to exploit these. How quickly it exploits or how long it explores is a trade-off that you have to make.

Hyperparameters for RL Optimization Algorithm

The screenshot below shows the training configuration for the final model. This includes the action space, which is where we set the maximum and minimum speeds for the agent, as well as the maximum and minimum steering angles. Additionally, the screenshot shows the settings for the hyperparameters, which were not altered for the final model. A separate model, using the same reward function but an increased gradient descent batch size, number of experience episodes between each policy-updating iteration, and decreased learning rate led to much more consistent training, but slower peformance upon evaluation.

Elements of Reward Function

Initializing parameters

List of Variables for the Reward Function

The following code block shows the first few lines for the reward function. Here, we give a brief summary of the reward function's incentives and penalties. We also read the input parameters to memory in order to use them throughout the rest of the function. After we read in the parameters, we create three variables that hold values that we will later use to incentivize staying near center line of the track. We created thresholds and initialized the reward value in the final part of this code block.

# Read input parameters
distance_from_center = params['distance_from_center']
track_width = params['track_width']
speed = params['speed']
waypoints = params['waypoints']
closest_waypoints = params['closest_waypoints']
progress = params['progress']
steps = params['steps']

# Calculate 3 marks that are increasingly farther away from the center line
marker_1 = 0.1 * track_width
marker_2 = 0.25 * track_width
marker_3 = 0.5 * track_width

# initialize thresholds and reward
DIFF_HEADING_THRESHOLD = 6
SPEED_THRESHOLD = 1.8
TOTAL_STEPS = 270      # roughly 15 steps per second, 18 sec default lap
reward = 5

Progress Incentive

After some research, we learned that the agent performs roughly 15 steps per second. The AWS DeepRacer Developer Guide (https://docs.aws.amazon.com/deepracer/latest/developerguide/deepracer-reward-function-input.html) provides information regarding all of the parameters that we can modify in the reward function. The guide also provides code that shows examples of the paramaters in use. One of these examples uses the "progress" and "steps" parameters in order to code a reward function that incentivizes the agent to make more progress more quickly. To write the code block below, we used the same logic, but modified the specific numbers.

############################
# Steps and progress, check every 30 steps. The less steps it takes to reach 100% progress means finished faster
############################

# reward less steps with greater progress if faster than 18 sec
if (steps % 20) == 0 and progress/100 > (steps/TOTAL_STEPS):
    reward += progress - (steps/TOTAL_STEPS)*100

Essentially, the "TOTAL_STEPS" variable is calculated using 15 steps per second * 18 seconds. 18 is an arbitrary value that we decided on based on the performance of the model using a default reward function. Since the agent performs 15 steps per second, it should complete 270 steps in 18 seconds. If the agent has made more progress around the track than it would have if it was driving at a constant 18 second pace, the agent is rewarded. This reward is calculated based on how much further ahead it is than it would've been if driving at an 18 second pace. We check every 20 steps.

Waypoints Incentive

The waypoints parameter is an ordered list of milestones along the track center. Each track has its own unique list of waypoints. Each milestone is described as a coordinate of (xw,i, yw,i). The image below helps to visualize the manner in which waypoints are placed along a track. It is possible to retrieve the coordinates of any desired milestone at any point. This fact can be used to determine a corner in the future.

In the code block below, we find the previous milestone, next milestone, and a milestone 6 points in the future. We then calculate the measure (in degrees) of the current direction we are facing, and the direction we will face 6 points in the future. We find the difference between these two variables. If this difference is greater than a certain value (stored in DIFF_HEADING_THRESHOLD), it indicates that a corner exists close ahead of the agent at the time. If the difference is greater than the threshold and the agent is going faster than the speed threshold (stored in SPEED_THRESHOLD), we penalize the agent. This works to incentivize the agent to take corners more slowly.

#############################
# Waypoints: referenced code from https://github.com/MatthewSuntup/DeepRacer/blob/master/reward/reward_final.py
#############################

# finding previous point, next point, and future point
prev_point = waypoints[closest_waypoints[0]]
next_point = waypoints[closest_waypoints[1]]
future_point = waypoints[min(len(waypoints) - 1, closest_waypoints[1]+6)]

# calculate headings to waypoints
heading_current = math.degrees(math.atan2(prev_point[1]-next_point[1], prev_point[0]-next_point[0]))
heading_future = math.degrees(math.atan2(prev_point[1]-future_point[1], prev_point[0]-future_point[0]))

# calculate difference between headings
# check we didn't choose reflex angle
diff_heading = abs(heading_current-heading_future)
if diff_heading > 180:
    diff_heading = 360 - diff_heading

# if diff_heading > than threshold indicates turn
# so when a turn is ahead (high diff_heading)
# penalize high speed, reward slow speed
if (diff_heading > DIFF_HEADING_THRESHOLD) and speed >= SPEED_THRESHOLD:
    reward -= 4

Centerline Incentive

The final part of the reward function was to incentivize the agent to stay near the centerline of the track. The following code is a modified version of the first default reward function given by the AWS DeepRacer console.

############################
# Center line incentives
############################

# give higher reward if the car is closer to center line and vice versa
if distance_from_center <= marker_1:
    reward += 5
elif distance_from_center <= marker_2:
    reward += 4
elif distance_from_center <= marker_3:
    reward += 3
else:
    reward -= 4

Reward Graph

The reward graph shows the model's progress as it trains. It is a line graph with three lines: Average reward (training), Average percentage completion (training), and Average percentage completion (evaluating). The final reward graph is shown below.

Testing Evaluation

After the training had completed, we evaluated the model using the same track that it trained on. The AWS Console allows us to perform multiple evaluations. Below, we show the result and a video of our best evaluation that we performed, where the agent drove three laps around the track during each evaluation.

Evaluation Video

final_model_Evaluation_2.mp4

Evaluation Results Table

image

Physical Testing

The entire goal of the AWS DeepRacer project is the train a Reinforcement Learning model that can be inputted into a vehicle in the real-world. To truly test our model we built the vehicle with all the necessary sensors to collect data and extract the parameters for running our trained model from earlier.

Building the Robot

Unboxed AWS DeepRacer Kit

image

Assembled Car

image

We completed the assembly of our AWS DeepRacer Car using the instructions on Amazon's Guide. After building the car we had to connect and calibrate it using a WiFi network to access the car's control panel for uploading and running our model. Below is a screenshot of the control panel for the robot.

image

Building the Track

re:Invent 2018 Track Template

image

With our time constraints and the high price and lack of availability of materials, we were not able to construct a full track but rather completed a subset of the track which was the left hand side (boxed in green above).

Final Track Subset

image

Evaluating the Model

Through the AWS DeepRacer control panel, we imported our best trained model that we had previously evaluated virtually (as shown earlier). Due to wifi limitations as well as the imperfect track conditions, we had to limit our vehicle's speed to ~50% which is why the run is not nearly as fast as the virtual. We can see in this video that the model seemed to be able to handle 110° and 78° turns as well as speed up on a straight-away.

Successful Physical Run

deepracer-video.mp4

Discussion

Throughout the development of our AWS DeepRacer, we experienced many challenges and moments of learning with relation to both the creation of a Reinforcement Learning model as well as the physical implementation and execution of projects. A huge consideration for the AWS DeepRacer project is that the 2 components (virtual and physical) can take months alone to set up and fine-tune for competitions so the limitation on time we had resulted in a lot less training and testing than ideal.

Physical Limitations

  • The cost of materials for a full track and time to construct it with the correct angles. Trying to make a subset of the re:Invent 2018 track with poster boards and tape was difficult because we lacked the tools and space. Since our model was trained on a virtual environment with no flaws in the track, all the small issues negatively affected the performance of our robot's real-life evaluation.
  • WiFi limitations were very frustating since it made connecting our laptop to the robot and streaming the camera's content sluggish and sometimes non-existent. The AWS DeepRacer logs also have no information on connecting to a network that is account secured like university WiFi. Our solution was to bring our own router to create a separate network that both the laptop and car could connect to.
  • The AWS DeepRacer Compute Module caused many issues that were nearly undiagnosable without previous knowledge. Alongside WiFi limitations were the challenges that were caused by a possibly corrupt ruleset for the compute module. The webserver would not open on our computers when connected to the car even after factory resetting by installing a fresh Ubuntu OS onto it and resetting the AWS credentials. The solution was to add an "http" rule to the compute module though that was nowhere to be found in any documentation or forums.

Virtual Limitations

  • Our free AWS accounts only had 10 hours for training & testing models which made it hard to play around with all the combinations of hyperparameters as well as the fine-tuning for the reward function values which may have improved our time. Our approach to mitigate this was having each person test different models on separate accounts to get the most efficiency.

Results

By the end of this project we successfully created a reward function for a Reinforcement Learning model and trained it to drive the AWS DeepRacer car around a track autonomously. Our virtual times of ~13 seconds for the 2018 competition track were not perfectly translated to physical testing for reasons as mentioned above but we still were able to get the model loaded onto the actual car and working.

The most valuable part of this was being able to get hands-on experience with Applied Reinforcement Learning with the autonomous driving in this project. The ability to take the theoretical concept of Markov Decision Process and dynamically programmed trial-and-error and create something practical provided insight into the real-world use cases of Machine Learning.

Considerations and Future Goals

While the learning process is crucial, it's even more important to question the implications of what you're doing. This is especially true as the project you are working on is more applied/practical. We dove deeper into the history of driving and the possible biases and impacts of our project in the real world. An example of the considerations that should be taken into account is how our project and trained models wouldn't be able to be applied in many real-world scenarios as roads are very different depending on where you are. This can cause a disparity in access to this new technology. For more in-depth discussion on these questions feel free to read the EthicalConsiderationsWksht.md within this repo.

If we had more time in the future, some goals that we would love to accomplish are the optimization of our model to be faster than 10 seconds per lap as well as testing more on other tracks to avoid overfitting to one track. With regards to the physical environment, we'd hope to create a more permanent track using EVA foam pieces so that it is collapsable/transportable. Both virtually and physically, entering a competition is something that would allow us to compare and discuss our models and training environments to others who may have more experience or understanding in the field of Reinforcement Learning and Autonomous Driving.

About

final-project-jcar-uark created by GitHub Classroom

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%