3D Path & Pose Viewer

The 2-D scene-video overlay shows what the camera sees. This is the 3-D companion: the drone’s path and pose in the world, driven by the same run.jsonl stream that vio_pipeline emits. Drag to orbit, scroll to zoom, and use the scrubber (or ▶) to step through the run.

▶ ⤢

ground truth estimate 3σ covariance camera frustum landmarks

What you are looking at

• Green line / triad — the ground-truth trajectory and the true pose at the current frame.
• Orange line / triad — the estimated trajectory and the filter’s current pose estimate.
• Yellow ellipsoid — the filter’s position-covariance at 3σ (≈ 97% per axis). This is the whole reason for the 3-D view: it makes the filter’s belief about its own uncertainty visible.
• Grey points — the static landmark cloud the camera observes (the synthetic scene’s true 3-D points).
• Cyan pyramid — the camera frustum at the estimated pose, reconstructed from the pose and the camera extrinsics/intrinsics. Watch it sweep across the cloud — the landmarks it covers are what make the motion observable.

Why the ellipsoid matters (the #212 story)

A trustworthy filter keeps the green ground-truth marker inside its covariance ellipsoid — the true pose lies within the uncertainty it reports. When the ellipsoid is too small and the green marker drifts outside it, the filter is over-confident: certain about a wrong position. The HUD flags this the moment the true error exceeds the 3σ ellipsoid.

This is exactly the over-confidence tracked in issue #212: on this sample run the reported position σ shrinks to ~13 cm while the true error grows past 2 m — the ellipsoid collapses to a fraction of the actual error. A controller trusting that covariance would act decisively on a wrong position. See the pipeline how-to and the canonical contract doc for the stage-by-stage decomposition.

Generate your own

cmake -B build -DBUILD_TARGET_KPU=OFF && cmake --build build -j$(nproc)
./build/tools/vio_pipeline --robot drone --out /tmp/run
# → /tmp/run/run.jsonl  (drop it in docs-site/public/data/ to view here)

Render it to an mp4 (offline)

The same viewer renders headlessly to a video — useful for talks and slides:

# one-shot: experiment → 3D video
scripts/vio_scene_video.sh --3d --robot drone -o /tmp/run    # → /tmp/run/scene3d.mp4

# or straight from an existing run.jsonl
node docs-site/scripts/gen-scene3d-video.mjs --data /tmp/run/run.jsonl --out /tmp/run/scene3d.mp4

Needs playwright (cd docs-site && npm i -D playwright && npx playwright install chromium) and ffmpeg. It drives this exact viewer frame-by-frame in a headless browser, so the video matches what you see above.

Each estimate record now carries the 3×3 position covariance block (pcov) that drives the ellipsoid:

{"type":"est","t":13.95,"q":[...],"p":[1.23,3.36,0.38],"pos_err":2.59,
 "pcov":[0.0165,-0.0016,0.00016, -0.0016,0.0206,0.0001, 0.00016,0.0001,0.0108]}