Carlos Valadares

Develop and enhance algorithms for Prediction and Planning blocks inside the Autonomous Driving pipeline using Deep Learning.

First year:

• Creation of a modular and interpretable system capable of producing a safe short-term trajectory, taking into account the surrounding environment of the car, traffic rules, and navigation goal.
• Publication of developed method on IEEE Intelligent Vehicles Symposium (IV): MIMP: Modular and Interpretable Motion Planning Framework for Safe Autonomous Driving in Complex Real-world Scenarios

Second year:

• Internship Tutor: Reinforcement Learning for autonomous driving in complex urban environments.
• Pioneer in the exploration and understanding of ML/AI pipeline, including Perception and Routing Neural network models after partner acquisition (AIMotive).
• Increase the speed of the trajectory prediction pipeline by 10%.
• Promotion to Lead Software Engineer.

Third year (in progress):

• Team lead: responsible for managing and distributing activities as well as follow up progress on the development of a new benchmark.
• Creation of an evaluation benchmark in order to detect critical trajectory failures using an in-house dataset.
• Hundreds of critical failures detected and reported successfully.

R&D For autonomous vehicles monocular percetion (2D/3D Object Detection, Semantic Segmentation,Depth estimation) at the DeepPerception project with an industrial partner.

• Transfer all network models and AI pipeline from Tensorflow to Pytorch
• Significant increase in Monocular 2D and 3D object detection performance on internal real-data evaluation dataset and Kitti benchmark.
• Change of the feature reference frame from Camera to Birds-Eye-View (BEV) using lift strategies

R&D of Reinforcement Learning agents in order to securely drive a car autonomously in urban scenarios.

• Use of CARLA simulator.
• Use of state-of-the-art off-policy reinforcement learning algorithm: Soft Actor Critic (SAC) plus known RL tricks to increase performance.
• The agent was able to navigate successfully in all scenarios of the CARLA Leaderboard Challenge (without dynamic agents).

Development of algorithms to extract optical index and localization of objects or cells from backpropagated holograms using convolutional neural networks.

• Creation of a Deep Learning pipeline from scratch, including construction of training/evaluation dataset, neural network structure, loss, logging, training, evaluation and more in order to quickly estimate optical index and localization
• Publication of research results on SPIE Photonics Europe conference: Spherical object segmentation in digital holographic microscopy by deep-learning