Results

PIA Days

PIA days are organized every semester to provide an opportunity for PIA students to present their work.

AI Driving Olympics 5th and 6th editions

In 2020 PIA students competed with 6 different solutions at the 5th edition of the AI Driving Olympics (AIDO) which was part of the 34th conference on Neural Information Processing Systems (NeurIPS). There was a total of 94 competitors with 1326 submitted solutions, thus we proudly announce that our team ranked top in 2 out of 3 challenges.

PIA students also competed in the 6th edition of AIDO with excellent results.

Research results

Imitation learning in DuckieTown: Imitation Learning uses the demonstrations of an expert to uncover the optimal policy and it is suitable for real-world robotics tasks as well. In this case, however, the training of the agent is carried out in a simulation environment due to safety, economic and time constraints. Later, the agent is applied in the real-life domain using sim-to-real methods. In our research, we applied Imitation Learning methods that solve a robotics task in a simulated environment and used transfer learning to apply these solutions in the real-world environment. Our task is set in the Duckietown environment, where the robotic agent has to follow the right lane based on the input images of a single forward-facing camera. We presented three Imitation Learning and two sim-to-real methods capable of achieving this task. A detailed comparison was provided on these techniques to highlight their advantages and disadvantages.

Publications:

Simulation to Real Domain Adaptation for Lane Segmentation: As the cost of labelling and collecting real world data remains an issue for companies, simulator training and transfer learning slowly evolved to be the foundation of many state-of the-art projects. In our research these methods were applied in the Duckietown setup where self-driving agents can be developed and tested. Our aim was to train a selected artificial neural network for right lane segmentation on simulator generated stream of images as a comparison baseline, then use domain adaptation to be more precise and stable in the real environment. We have tested and compared four knowledge transfer methods that included domain transformation using CycleGAN and semi-supervised domain adaptation via Minimax Entropy. As the latter was previously untested in semantic segmentation according to our best knowledge, we have contributed to showing it is indeed possible and produces promising results. Finally we have shown that it could also create a model that fulfills our performance requirements of stability and accuracy. We also showed that the selected methods are equally eligible for the simulation to real transfer learning problem, and that the simplest method delivers the best performance.

Publication:

Efficient Neural Network Pruning Using Model-Based Reinforcement Learning: Model compression plays a vital role in the deployment of neural networks (NNs) in resource-constrained devices. Rule-based conventional NN pruning is sub-optimal due to the enormous design space that cannot be examined entirely by hand. To overcome this issue, automated NN pruning leverages a reinforcement learning agent to automatically find the best combination of parameters to be removed from a given model. We proposed a novel RL-based automated pruning algorithm that, unlike existing RL-based methods, determines the environmental variables using a State Predictor Network as a simulated environment instead of validating the pruned model in run time. Testing our method on the YOLOv4 detector, a model with 49 % sparsity was produced with 7.2 % higher mAP. This result outperforms our handcrafted pruning methods for YOLOv4 by 2.3 % mAP and 17.1 % sparsity. Regarding total development time, our method is 146.2 times faster than the state-of-the-art PuRL method using NVIDIA Titan X GPU. The implementation of the proposed solution is available at: https://github.com/bencsikb/Efficient_RLPruning

Publications:

Semantically Consistent Sim-to-Real Image Translation with Neural Networks: Texture-swapping of images has industrial benefits besides artistic stylization and photo editing, e.g. simulated images could be modified to look like real ones to train Computer Vision methods. Autonomous driving research could largely benefit from this as its neural network-based perception systems need a large amount of labeled training data. However, the sim-to-real texture swapping is a demanding challenge because of the large gap between the two domains. Another requirement is that the semantic meaning of the photo should not change during the translation. In our experiments, we found that SOTA algorithms struggle with these expectations, so in our work, we improved a former method by taking advantage of the semantic labeling of the training datasets. We showed that with our two improvements, we can better conserve the scene of the image during the sim-to-real translation while the photorealism of the output image does not significantly change.

Publications:

Sim-to-real reinforcement learning applied to end-to-end vehicle control: In our work, we studied vision-based end-to-end reinforcement learning on vehicle control problems, such as lane following and collision avoidance. Our controller policy was able to control a small-scale robot to follow the right-hand lane of a real two-lane road, while its training was solely carried out in a simulation. Our model, realized by a simple, convolutional network, only relies on images of a forward-facing monocular camera and generates continuous actions that directly control the vehicle. To train this policy we used Proximal Policy Optimization, and to achieve the generalization capability required for real performance we used domain randomization. We carried out thorough analysis of the trained policy, by measuring multiple performance metrics and comparing these to baselines that rely on other methods. To assess the quality of the simulation-to-reality transfer learning process and the performance of the controller in the real world, we measured simple metrics on a real track and compared these with results from a matching simulation. Further analysis was carried out by visualizing salient object maps.

Publication: