TY - GEN
T1 - Intelligent Robot Learning Research – Bridging the Gap between Academia and Industry
AU - Lynnerup, Nicolai Anton
PY - 2022/3/16
Y1 - 2022/3/16
N2 - Reproducing the work of others is one of the main pillars of scientific research. It allows researchers collectively to advance their research fields and cross-validate results across institutions, circumstances, experimental variation etc. When research is irreproducible, or unable to be replicated, it hinders technological uptake in industry regardless of research area and furtherimpairs the credibility of the academic field. For real-world robot learning reproducibility issues have resulted in a crisis where research is published – even in large, well-known and respected conferences and journals – despite many studies failing to report all necessary details accurately. We refer to this kind of research as “look ma no hands” research.This thesis investigates technologies currently feasible and applicable for realworld – that is, industrially applicable – robot learning, with a focus on reinforcement learning (RL). We present a set of methodologies and guidelines that, combined, ensure the reproducibility of the computational research. We consider the complete development pipeline, from idea to deployment, and include important aspects such as data management plans, safety and infrastructure for orchestrating the learning pipeline. These methodologies are derived from four real-world cases, both from academic and industrial settings. We argue that such an approach is essential not only for scientific progress but also for engendering the trust required for industry to commit investment in the results of research in this area.
AB - Reproducing the work of others is one of the main pillars of scientific research. It allows researchers collectively to advance their research fields and cross-validate results across institutions, circumstances, experimental variation etc. When research is irreproducible, or unable to be replicated, it hinders technological uptake in industry regardless of research area and furtherimpairs the credibility of the academic field. For real-world robot learning reproducibility issues have resulted in a crisis where research is published – even in large, well-known and respected conferences and journals – despite many studies failing to report all necessary details accurately. We refer to this kind of research as “look ma no hands” research.This thesis investigates technologies currently feasible and applicable for realworld – that is, industrially applicable – robot learning, with a focus on reinforcement learning (RL). We present a set of methodologies and guidelines that, combined, ensure the reproducibility of the computational research. We consider the complete development pipeline, from idea to deployment, and include important aspects such as data management plans, safety and infrastructure for orchestrating the learning pipeline. These methodologies are derived from four real-world cases, both from academic and industrial settings. We argue that such an approach is essential not only for scientific progress but also for engendering the trust required for industry to commit investment in the results of research in this area.
U2 - 10.21996/sxcb-3x60
DO - 10.21996/sxcb-3x60
M3 - Ph.D. thesis
PB - Syddansk Universitet. Det Tekniske Fakultet
ER -