Modes and Mechanisms in Reinforcement Learning for Robotics
Title: Modes and Mechanisms in Reinforcement Learning for Robotics
DNr: NAISS 2023/22-275
Project Type: NAISS Small Compute
Principal Investigator: Alexander Dürr <alexander.durr@cs.lth.se>
Affiliation: Lunds universitet
Duration: 2023-03-02 – 2024-04-01
Classification: 10207
Homepage: https://portal.research.lu.se/portal/en/projects/reinforcement-learning-in-continuous-spaces-with-interactively-acquired-knowledgebased-models(6bcfa32c-e9c7-468a-9acb-f433ad98a06a).html
Keywords:

Abstract

Our research field is Reinforcement Learning (RL) for robotics with different modes (e.g. language, sound, ...) and mechanisms (e.g. attention, curiosity, ...) and {task, skill, visual, knowledge} representations (e.g. embeddings, triples, ontology, graphs, ...). We make use of Neural Architecture Search (NAS) of medium to large Neural Networks (e.g. CNN, Transformer, ...). To benchmark, we evaluate our approach against implementations of methods from other authors to determine and compare the performance. We plan to apply explainability algorithms (e.g. Grad-CAM, RISE, SHAP, ... ) to check for plausibility of the agent's actions and reasoning.