[Closed] Master Internship on Meta and Transfer Learning for Deep Reinforcement Learning in Robotics

Topic: The internship is part of our research project into Deep Reinforcement Learning (DRL) [1] for the European SPRING project. SPRING aims to develop control mechanisms for mobile robots that will be employed in hospitals and health care environments. The robots should communicate with elderly people, their families and caretakers to inform, aid and entertain them.

We are developing DRL methods in this context to learn the robot’s movement behaviors, such as human aware navigation, how to best approach people or how to collect sensor information (visual and auditory) to allow successful conversations between the robot and humans. DRL utilizes deep neural networks [2] to represent its important components, i.e. the value function and policy.
A problem with Deep Learning and DRL is that they need many data samples to learn appropriate behaviors. This is especially a problem in robotics as it is time-consuming to collect data with a robot. We focus therefore on Meta-Learning [3] and Transfer Learning [4] methods that allow a fast adaptation of the learning and behaviors from already learned tasks to new ones. This can be used, for example, to first learn human aware navigation in a simulation and then transfer this knowledge to the physical robot to perform the same task without the need for relearning from scratch [5]. We are searching new and innovative learning mechanisms and deep network architectures to improve reinforcement learning in this context.

Task: During your internship, you will be reviewing relevant DRL approaches under the guidance of your supervisor. Together you will develop ideas for new architectures and learning mechanisms. Your task will then be to implement one method and to test it on a robotics task (either simulated or real). You will have to compare it to existing approaches and report the experimental results in a comprehensive manner in form of a small research paper.

Environment: You will be joining the Perception team, an international team of researchers and students at Inria Grenoble. The team has a strong background in audio-visual computation and its application to robotics. The team is headed by Patrice Horaud (Team leader) and Xavier Alameda-Pineda (SPRING Project Leader). You will be supervised by Chris Reinke (Postdoc) and Xavier during your internship.

[1] Li, Y. (2017). Deep reinforcement learning: An overview. arXiv preprint arXiv:1701.07274.
[2] Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press.
[3] Hospedales, T., Antoniou, A., Micaelli, P., & Storkey, A. (2020). Meta-learning in neural networks: A survey. arXiv preprint arXiv:2004.05439.
[4] Taylor, M. E., & Stone, P. (2009). Transfer learning for reinforcement learning domains: A survey. Journal of Machine Learning Research, 10(7).
[5] Kang, K., Belkhale, S., Kahn, G., Abbeel, P., & Levine, S. (2019, May). Generalization through simulation: Integrating simulated and real data into deep reinforcement learning for vision-based autonomous flight. In 2019 (ICRA) (pp. 6008-6014). IEEE.

Comments are closed.