When the ego decision changes.
Windblown hat · one hat retrieval · overlapping 3- and 4-agent interactions · earlier cues, left turns and following gaps · ends by +20 s
Loading…
Decision history
Ego meta-action · same world time
STOP / DECELERATE / KEEP_SPEED / ACCELERATE describe the following 4 seconds of each replay; current behavior is shown below each label. Lateral action is FOLLOW_LANE. These are replay annotations, not model predictions.
Faint paths and ghost positions show alternatives at the same world time. Shared paths are merged for clarity. Ghosts are position guides; the selected leaf shows full motion. Drag the time slider or press Play.
The amber line is a rehearsal stop reference just before the mapped intersection entrance (the map has no explicit stop-line layer). Dashed bars show vehicle front stopping positions; rings show planned centre or waiting points. Colours match the agents. These review overlays are excluded from clean GLBs and image-generation frames.
Rehearsal checks
Compare ego responses
The same causal policy is used across the tree. Candidate comparisons measure geometric safety, braking, and progress at the rehearsal end. They do not establish global optimality.
| Response | Person gap | Rear gap | Peak brake | Peak jerk | Progress at the rehearsal end |
|---|
Assumptions and review evidence
Scenario definitions and numeric checks · Visual review record · Meta-action comparison evidence