Interactive 4D reconstruction
Drag to rotate; scroll to zoom; Shift + drag to pan. Coordinates are in the robot base frame (m). X: red, Y: green, Z: blue. Playback is an inspection rate, not wall-clock time.
Reconstruction and temporal alignment
Predicted RGB and decoded depth are back-projected using the calibrated camera model. Observed clouds use recorded RGB and metric depth. Each point cloud is reconstructed independently: these are visible-surface sequences, not persistent point trajectories. Unobserved surfaces are not completed.
Comparison mode uses prediction sampling times and the final observation. Only exact action-index matches are displayed; missing frames remain empty. The observation stream includes every recorded state. The prediction stream concatenates successive prediction rounds conditioned on the available observation history.
Invalid depth and values outside (0.01, 4) m are excluded. The optional prediction filter removes depth-code residuals above 0.02. For delivery, depth is quantized to 1 mm; this display compression does not affect the flow-tracking metrics. RGB is retained at native resolution. Workspace crop uses X=[0.1,1.1], Y=[−0.6,0.6], Z=[−0.08,0.9] m.