Experiments need to be planned far in advance. Ege prepared a document as discussion basis. Two major experiments on investing the effects of transfer learning and on shared control.
– First experiments could be implemented using a smartphone app. Key requirements are that
1. the trajectories (e.g., of objects or mouse cursors) are recorded with a sufficient time resolution.
2. The tasks are modulated to investigate the learning of a new skill (e.g., change the target location or the object size or the object weight).
3. The learning needs to be sufficiently complex to identify an effect of model transfer. E.g., a task should be learned in at least 5-10 trials.
– First models could be based on probabilistic movement primitives where we transfer the main weights or all parameters as done in https://ai-lab.science/wp/IROS2019Stark.pdf