PIA project’s perseverance
Author: Robert Moni
Sorry for the clickbait, but no, we aren’t the engineers that put together the Perseverance Mars rover, although we do possess the right amount of perseverance to achieve our goals.
This blog post attempts to provide a summary of our latest research achievements under the aegis of the PIA project.
PIA Project
The Professional Intelligence for Automotive Project was founded in 2019 as a cooperation project between Continental Deep Learning Competence Center Budapest and the SmartLabs of the Department of Telecommunications and Media Informatics (TMIT) and the Department of Control Engineering and Information Technology (IIT) at the Budapest University of Technology and Economics.
The motivation of the cooperation project is to establish a strong relationship between the industry and students. The goal is to support research and development activities for BSc, MSc and PhD students in the field of Artificial Intelligence and automotive. The participating students benefit from professional support offered by the Deep Learning engineers and expert of Continental, and financial support in the form of scholarship, also internship and future job opportunities are obtainable.
During 2020 there were 10 students conducting research and development activities. We published 6 conference papers and 6 of us competed in the 5th AI Driving Olympics competition where we gained four 1st, four 2nd and three 3rd placements in the self-driving challenges (simulation and real environments considered).
A brief explanation of our research field and our achievements presented in Medium blog posts is summarized in the following.
Research field
Our research focuses on the sphere we get by the conjunction of Deep Learning and Autonomous Driving. More precisely we attempt to elaborate on new approaches for Vision-based Environment Perception and Policy-based Vehicle Control. For environment perception, we applied Supervised Learning and Unsupervised Learning methods, while for vehicle control, Reinforcement Learning methods were elaborated.
Vision-based Environment Perception
Environment Perception refers to the active monitoring of the surroundings of a vehicle to extract meaningful task-related information for safety and control. The main sensors responsible for sensing are the Camera, Radar, Lidar and Communication devices (e.g. V2X via Wi-Fi or 5G) depicted by the following image.
We focus on Camera-based Environment Perception and aim to provide state-of-the-art solutions for the following applications:
- Semantic and Instance segmentation
- Deep Learning-based Semantic Segmentation in Simulation and Real-World for Autonomous Vehicles
- Deep Learning-based Semantic Segmentation in Simulation and Real-World for Autonomous Vehicles — part 2
- Deep Reinforcement Learning for Vehicle Control based on Segmentation *only the first part
- Apply semi-supervised learning for semantic segmentation
2. Depth estimation
3. Multi-task Learning
- Object Detection on CityScapes Dataset
- Disjoint Datasets in Multi-task Learning with Deep Neural Networks for Autonomous Driving
Policy-based Vehicle Control
Deep neural networks possess the ability to approximate complex functions. Self-driving based on camera input is indeed a complex function which we attempt to approximate with Deep Reinforcement Learning (DRL) methods. The aim is to reach an optimal or near-optimal self-driving policy represented by the weights in the deep neural network.
There are several ways to achieve such a self-driving policy. We developed multiple solutions considering end-to-end or modular architecture designs, on-policy or off-policy DRL algorithms and even Imitation Learning.
- End-to-end DRL-based self-driving agent
End-to-end design refers to the architectural design choice, where the agents learn directly from raw data.
- Controlling Self-Driving Robots with Deep Reinforcement Learning
- Lane following in the Duckietown environment under extreme conditions
- Applying transfer learning to the autonomous driving task
2. DRL-based self-driving agent with a modular design
Modular design refers to the architectural design choice, where the agents learn from extracted observations, usually, low-dimensional representations provided by a separate deep neural network with the scope to extract meaningful task-related information from the raw data (e.g. Vision-based Environment Perception or Representation Learning).
- Deep Reinforcement Learning for Vehicle Control based on Segmentation
*second part - Supervised and Unsupervised Representation Learning for Reinforcement Learning
3. Imitation Learning
- Imitation Learning in the Duckietown environment
- Using Transfer Learning to solve the simulator-to-real problem in the Duckietown environment
Publication and conferences
Other than medium blog posts, we did contribute officially to science via articles:
- P. Almási, R. Moni and B. Gyires-Tóth, “Robust Reinforcement Learning-based Autonomous Driving Agent for Simulation and Real World”, 2020 International Joint Conference on Neural Networks (IJCNN), Glasgow, UK, 2020, pp. 1–8, doi: 10.1109/IJCNN48605.2020.9207497.
- M. Tim, M. Szemenyei and R. Moni, “Simulation to Real Domain Adaptation for Lane Segmentation”, 2020 23rd International Symposium on Measurement and Control in Robotics (ISMCR), Budapest, Hungary, pp. 1–6, doi: 10.1109/ISMCR51255.2020.9263406.
- A. Kalapos, C. Gór, R. Moni and I. Harmati, “Sim-to-real reinforcement learning applied to end-to-end vehicle control”, 2020 23rd International Symposium on Measurement and Control in Robotics (ISMCR), Budapest, Hungary, pp. 1–6, doi: 10.1109/ISMCR51255.2020.9263751.
- B. Andras, R. Moni, B. Gyires-Tóth, “Sample-Efficient Deep Reinforcement Learning with Visual Domain Randomization”, 2020 Students’ Scientific Conference (TDK), Budapest University of Technology and Economics
- L. Zoltán, R. Moni, M. Szemenyei, “Imitation Learning in the Duckietown environment”, 2020 Students’ Scientific Conference (TDK), Budapest University of Technology and Economics
- P. Almási, B. Gyires-Tóth, “Real-world autonomous driving using deep reinforcement learning and domain randomization”, 2020 Students’ Scientific Conference (TDK), Budapest University of Technology and Economics
Competitions
Last but not least we are proud of our winning entries at AI Driving Olympics 5. A detailed blog post about our solutions can be found here: PIA project achievements at AIDO5.
