Improve online execution v2

Hey everyone,

another issue to adress the current limitations and problems of the anomaly detection.

Based on this discussion and a physical meeting the next steps are the following:

• train with old data (reduce quantity of bags)
• train with more epochs (new data), to overfit and be perhaps more sensitive for anomalies
• include EE pose and velocity
• train a predictive model
• implement a different architecture (Niki)
• literature review (What have others done? architecture, window_size, amount of data, tricks)
• (no prio) quantitative metric for fast evaluation (additionally to the visual validation)
• (no prio) easy labeling improved evaluation (Felipe)