What I did
- Inserted markers into log files. [what were the markers?]
- Clipped each trial by barrier. Trajectories begin and end at dial-stop markers.
- Aligned trajectories at the [???] marker. The position and time of the marker defined the point of 0-time and the 0-arclength.
- Projected down into 2-d.
- Computed per-sample measures: distance-along-arc, frenet frame, velocity/accel/jerk relative to the frenet frame, path curvature.
- PCA based on pointwise similarity of scalar per-sample measures. [some have been done. some yet to do] [cropping or stretching to align?]
- Defined and computed several per-barrier measures (i.e. “dependent variables”) to describe each trajectory. [list these]
- Exposure. [see Jess’s writeup]
What we were trying to find.
Where we went wrong
- Measures taken were able to distinguish the conditions reasonably well (especially the artificial conditions). But we were really interested in success or failure in terms of exposure. People were exposed a lot in WIP and Joystick, and not at all in VR/Cowl/Real. So any per-barrier measures that allow us to distinguish WIP and Joystick from the walking conditions will also correlate with exposure. So we’re not finding measures that predict performance so much as we’re finding measures that distinguish artificial walking.
- It is obvious to all of us now that the don’t-be-exposed task was not difficult enough to sufficiently distinguish
- Walking in place was not good enough. Even if someone could have eventually become good enought at WIP to do a decent job, there was not nearly enough training time to establish any sort of learning plateau. Given only a few minutes practice at most, the subject was still on a steep part of the learning curve. [therefore we should see improvement between trials 1 and 3 of the WIP cases. this is something we could test if we wanted to.]
- The joystick did not allow strafing, so you must move in the direction you look. Observing real-walking cases, people spent much of their time looking and moving in different directions. Thus the user loses a degree of freedom that he/she would otherwise want to be using.
- Any results we claim about the effectiveness of Joystick and WIP will pertain only to our particular implementations of these interfaces (or implementations that are very near to them). Neither of these implementations is anywhere near to state-of-the-art. Anyone who is serious about training would pick better implementations than what we had, and therefore our results would not be of any help to them.