Safely learning to help safely

Activities of daily living (ADL) represent a collection of abilities required to be able to independently care for oneself [1]. These include activities such as feeding, ambulating and continence [2].

Bildet kan inneholde: grønn, maskin, personlig verneutstyr, teknologi, vitenskap.

This master project, which is related to the newly founded PIRC project (https://www.uio.no/ritmo/english/projects/pirc/), aims at aiding people living with limited and/or decreasing autonomy via a robot assistant to perform ADL. An important aspect of the project is that the robot assistant must be able to assist the user in an autonomous and robust manner.

Reinforcement learning (RL) algorithms enable robotic systems to robustly learn high dimensional control policies. To achieve these results, RL algorithms assume the ability of the agent to freely explore and interact with the environment. However, in scenarios within healthcare, exploration often comes with unacceptable risks making traditional RL algorithms inapplicable. In contrast, offline RL [3] proposes to learn an agent’s policy by relying solely on already collected trajectories, thus removing the need to perform active (and dangerous) exploration in the real world.

Thus, this master project proposes to explore how to safely learn a policy for the robot assistant (either the mobile robot Tiago (https://pal-robotics.com/robots/tiago/) or a robotic arm (https://www.universal-robots.com/products/ur5-robot/)) to interact in close collaboration with their user to improve their quality of life. 

Note that other research areas could be explored instead of (or as a complement) to Offline RL (e.g., Safe RL [4], Causal RL [5]).

 

References

[1] Edemekong, Peter F., Deb L. Bomgaars, and Shoshana B. Levy. "Activities of daily living (ADLs)." (2017).

[2] Wiener, Joshua M., et al. "Measuring the activities of daily living: Comparisons across national surveys." Journal of gerontology 45.6 (1990): S229-S237.

[3] Levine, Sergey, et al. "Offline reinforcement learning: Tutorial, review, and perspectives on open problems." arXiv preprint arXiv:2005.01643 (2020).

[4] Garcıa, J. and Fernández, F., 2015. A comprehensive survey on safe reinforcement learning. Journal of Machine Learning Research, 16(1), pp.1437-1480.

[5] Bareinboim, Elias. Causal Reinforcement Learning, ICML Tutorial [https://crl.causalai.net/]

Publisert 18. okt. 2021 14:21 - Sist endret 18. okt. 2021 14:21

Veileder(e)

Omfang (studiepoeng)

60