Speakers: Rituraj Kaushik
Robotics Seminar Series. Eighth Session – 14th August 2020, 15:00-16:00, via zoom. Link to event: https://aalto.zoom.us/j/238082076
As soon as the robots step out in the real and uncertain world, they have to adapt to various unanticipated situations by acquiring new skills as quickly as possible. Unfortunately, on robots, current state-of-the-art reinforcement learning algorithms (e.g., deep-reinforcement learning) require prohibitively large interaction time to train a new skill. In this talk, we explore methods to allow a robot to acquire new skills through trial-and-error within a few minutes of physical interaction. More precisely, we elaborate on how model-based robot learning can be used in conjunction with the priors from simulators to allow relatively complex robots, such as a hexapod, to adapt to broken legs and fault in the sensors within a minute of interaction and accomplish the task.
- Multi-objective Model-based Policy Search for Data-efficient Learning with Sparse Rewards (https://arxiv.org/pdf/1806.09351.pdf)
- Adaptive Prior Selection for Repertoire-Based Online Adaptation in Robotics (https://www.frontiersin.org/articles/10.3389/frobt.2019.00151/full)
- Fast Online Adaptation in Robotics through Meta-Learning Embeddings of Simulated Priors (https://arxiv.org/pdf/2003.04663.pdf)
- Black-Box Data-efficient Policy Search for Robotics (https://arxiv.org/pdf/1703.07261.pdf)