Experiment 3 recordings
SEAN Warehouse simulation · Jackal robot · Recordings from October 18, 2026
General navigation runs
Four development scenarios, each with seed 0. All runs are shown in full.
These runs use YOLO11 nano, the smallest detector, with four onboard RGB-D cameras (color and depth). The outside and overhead cameras are only for viewing; they do not provide input to the controller.
Controller input comparisons
Four encounters selected after inspecting E3 results. These are diagnostic examples, not a representative performance sample.
Baseline, YOLO and perfect-input results
Green cells reached the goal without contact. All green cells are tied on that outcome.
Swipe the table sideways to see all four encounters →
| Loading recorded outcomes… |
Perfect input supplies exact current positions and velocities to the same controller; it does not guarantee the best outcome. These four selected encounters do not establish an overall winner. Contact times are when the run ended, not goal-completion times.
Run observations
Comparison setup
The controller and initial scenario seed are shared. Later robot and pedestrian trajectories can differ as they respond to each other. Exact state is a diagnostic input, not a guarantee of the best outcome.
Clearance measurements and replay checks for this encounter
| Person information | Outcome | End time | Closest body clearance |
|---|
Method and recording details
Expand a setting for details. These settings are shared unless a condition explicitly changes the person input.
Onboard camerasFour RGB-D views · 640 × 480
Four simulated cameras face forward, left, rear and right, each with a 100° horizontal field of view. Forward conditions use only the front camera; surround conditions use all four. RGB-D provides color and depth.
Person estimatesYOLO11 · confidence ≥ 0.25
YOLO11 runs on color images. Detections below confidence 0.25 are discarded. Depth places detections in the ground plane, and a shared tracker estimates person positions and velocities. Nano is the smallest detector; extra-large is the largest.
ControllerSame controller · up to 0.8 m/s
The local controller evaluates forward speeds and turn rates over a two-second horizon, predicting people at constant velocity. It can brake or attempt recovery when forward motion is blocked. Robot pose, route and static walls are supplied in every condition.
Timing10 Hz updates · 0.1 s input delay
Cameras and control update every 0.1 s. Detector and original position-oracle observations have a fixed 0.1 s delivery delay. These comparisons equalize timing, not computational cost. Perfect current input bypasses that delay and the tracker.
Outside viewsViewer only · not controller input
General-run outside and overhead views were recorded during execution. Diagnostic overheads replay saved commands, with robot and pedestrian positions checked against the original logs. Blue marks the robot and red marks the pedestrian. Viewer cameras never feed the controller.
Scope of these recordingsSelected examples · SEAN social forces
General runs are development examples. Diagnostic encounters were selected after inspecting E3 and do not replace the full experiment. Outcomes measure contact and goal reaching, not human comfort. The pedestrian model is SEAN social forces; ORCA has not been evaluated here.