The project starts from results obtained on AI-aided inertial navigation, based on a classical training/testing procedure. Its aim is going to an on-line version of these navigation methods, which requires solving difficult issues such as deciding if the learned processing is still trustable or not, and re-learn rapidly when necessary.
Safran is an international high-technology group, operating in the aircraft propulsion and equipment, space and defense markets. Comprising a number of companies, Safran holds, alone or in partnership, World or European leadership positions in its markets. Within Safran Tech, Safran’s Corporate R&T Center in charge of developing disrupting and differentiating technologies for future products, the candidate will join the Information and Signal Technology department. We work on applying advanced mathematical and artificial intelligence algorithms to a wide range of applications (autonomous vehicles, non-destructive testing, design & manufacturing, health monitoring). We connect with startups, established companies and academic institutions to power our projects and benefit from the global research ecosystem. Within dedicated labs, we develop and operate prototypes at representative scale to be able to quickly test, iterate and demonstrate new concepts. Our team is growing to address the challenges of advanced control of autonomous vehicles. Context: Safran Tech and Mines Paristech started in 2018 a collaboration on new research aiming at enhancing inertial measurements units manufactured by Safran group with neural networks extracting kinematics information from streams of measurements acquired by various sensors. The project, at the boundary of classical navigation methods and deep learning, leverages the expertise of inertial navigation experts present in the Safran Tech machine learning team, a collaboration which quickly led to good results presented in several publications (1) (2) (3). Topic: The subject proposed for the postdoctoral contract is based on encouraging results already obtained on improving inertial navigation with deep learning, but is intended to address the specific difficulties raised by on-line adaptation. Currently, processing (either classical or based on machine learning) applied to a measurement stream is learned once and for all on a data set representative of the use cases of the navigation system. Between this learning phase and the implementation on the final product, engineers can check the results fit the performance requirements. This off-line learning strategy, usual in machine learning, suffers from two major drawbacks: - The use case and operating environment of the system have to be precisely known in advance - Sensor aging and failures cannot be taken into account These points motivate working on on-line adaptation of the processing over the product’s lifetime, without human intervention, based on sensor redundancy (IMU-GPS fusion, for instance, provides a ground truth usable for on-line fine tuning of the odometer measurement processing). If a network continues learning during operation, then a kind of monitoring is necessary, consisting in: - Deciding if the outputs of the network can be trusted or if a fail-safe mode has to be used, - Adapting or resetting the learning process when a model change is detected. The issue to be addressed by the candidate is determining how to make these decisions and on what basis. At least two sources of information will be considered: - Detection of out-of-distribution samples: quantifying epistemic uncertainty, - Statistical inconsistence: the outputs of the network are not compatible with data from other sensors. In a nutshell, the task is developing tools for the monitoring of a neural network during on-line learning, with AI-aided inertial navigation as main use case. Expected work: The candidate is expected to propose different approaches and identify existing work of the literature the topic can be related to, including but not limited to active and continuous learning (4) (5) some parts of reinforcement learning (6) (7) and recent research on uncertainty computation in neural networks (8). The candidate will have to determine if the problem can be addressed combining existing methods or requires an adapted theory. He/she will propose different solutions and test them either on public datasets or on data provided by Safran Tech. Depending on the obtained results, real experiments will be performed on a self-driving vehicle in a secure environment. It is expected that this work will lead to publications in top ranking conferences and journals. The candidate will be assisted in this task by experts of inertial navigation and sensor fusions who will provide him with the data, expertise and physical insight required by the mission.
Candidate profile: PhD in machine learning, with a specialization in on-line learning and/or uncertainty of neural networks
Bibliography: 1. Learning wheel odometry and IMU errors forlocalization. Brossard, Martin and Bonnabel, Silvère. 2019. International Conference on Robotics and Automation. 2. AI-IMU Dead Reckoning. Brossard, Martin, Barrau, Axel and Bonnabel, Silvère. 2019. Arxiv preprint. 3. RINS-W: Robust Inertial Navigation System on Wheels. Brossard, Martin, Barrau, Axel and Bonnabel, Silvère. Arxiv preprint. 4. Fine-tuning deep neural networks in continuous learning scenarios. Käding, C., et al. s.l. : Springer, 2016. Asian Conference on Computer Vision. 5. Active Learning from Demonstration for Robust Autonomous navigation. Silver, D., Bagnell, J. A. and Stentz, A. 2012. IEEE International Conference on Robotics and Automation. 6. Learning to adapt in dynamic, real-world environments through meta-reinforcement learning. Nagabandi, Anusha, et al. 7. Curious Meta-Controller: Adaptive Alternation between Model-Based and Model-Free Control in Deep Reinforcement Learning. Hafez,, Muhammad Burhan, et al. 8. Kendall, A. and Gal, Y. What uncertainties do we need in Bayesian deep learning for computer vision ? Advances in neural information processing systems. 2017.
(c) GdR 720 ISIS - CNRS - 2011-2019.