Trajectory Segment — supplies the comparison unit for → Preference Elicitation Protocol

The preference elicitation protocol never shows a human a full episode or a single state; it always presents two trajectory segments, rendered as short video clips between 1 and 2 seconds long (deep-rl-human-prefs, §"2.2.2 Preference Elicitation", p. 5). Because the paper cannot reset its high-dimensional environments to an arbitrary initial state the way Wilson et al. (2012) could, the two segments in a pair generally begin from different states, which complicates how a human is meant to interpret a preference between them (deep-rl-human-prefs, §"2.1 Setting and Goal", p. 4). The protocol's three-way response format, a preference, a tie, or 'incomparable,' exists to let a rater register that ambiguity rather than force a guess: ties become a uniform distribution over the two segments and incomparable judgments are simply dropped from the database. So the trajectory segment's job is not just to be short enough to query cheaply, but to be self-contained enough that a rater can still form a preference despite the two clips having no common starting point.