The collection app source, the data-processing pipeline, the trained model weights, and the privacy-processed dataset are not public yet. All four will be released once the review process concludes.
The collection app source, the data-processing pipeline, the trained model weights, and the privacy-processed dataset are not public yet. All four will be released once the review process concludes.
Collecting real-world navigation data for mobile robots typically requires platform-specific teleoperation, making large-scale collection expensive and difficult to scale. We introduce Universal Navigation Interface (UNI), a robot-free data collection paradigm that uses a four-wheeled rollator walker (rollator) and smartphone to collect physically constrained human demonstrations. Because the rollator cannot climb stairs, negotiate uncut curbs, or pass through narrow gaps, demonstrations are naturally biased toward wheeled-feasible routes. Using UNI, we collect 37.2 km of real-world navigation data and recover metric trajectories that directly supervise goal-conditioned navigation models. Fine-tuning visual-navigation models on UNI reduces trajectory prediction error by 17.4–24.8% on held-out UNI demonstrations. Evaluation on other navigation datasets shows benefits that vary by dataset and metric. We further demonstrate closed-loop transfer to a powered wheelchair in curb, staircase, and curb-cut scenarios. These results support low-cost physical proxies as a practical source of navigation supervision collected without the target robot.
Sessions are recorded on SensorVault straight to MCAP. That is the format that comes off the device, so a recording is readable by any MCAP viewer as-is — Foxglove above. Camera, LiDAR depth, IMU, GPS, audio, odometry and transforms play back together on one clock with no conversion step, no post-processing pass, and nothing to time-sync after the fact.
The collection rig is an off-the-shelf four-wheeled rollator walker with a standard handlebar phone mount. It takes about 30 minutes to assemble, costs roughly $250 excluding the smartphone, and needs no 3D-printed parts, custom-fabricated hardware, or platform-specific instrumentation. The operator pushes it with all four wheels on the ground, favouring ramps and curb cuts over stairs and uncut curbs, while its width constrains passage through narrow gaps. Full bill of materials below.
A custom iOS app records calibrated multimodal sensor streams on a LiDAR-equipped iPhone, so no separate sensors have to be integrated or synchronized. Any LiDAR-equipped iPhone works; an iPhone 16 Pro was used for consistency across all collection sessions. All streams share a monotonic clock and are stored in MCAP; RGB, depth, and VIO poses are synchronized at the ARKit-frame level. Focal lengths and principal-point coordinates are stored with each recording, and fixed focus keeps camera calibration consistent.
| Stream | Format | Rate |
|---|---|---|
| RGB | JPEG, 1280×720 | 30 Hz |
| Depth + confidence | Temporally smoothed ARKit LiDAR, 256×192 | 30 Hz |
| VIO pose | 6-DoF, per-recording origin | 30 Hz |
| Inertial | Accelerometer + gyroscope | 100 Hz |
| GPS | Device fixes | — |
| Audio | Continuous | — |
Low-confidence and out-of-range depth measurements are marked invalid before storage. Device VIO is recorded for reference but is not the released trajectory source: across the recorded VIO chunks spanning 10.8 h, 34.9% produced motion inconsistent with plausible human walking.
RGB recordings are reconstructed offline with Depth Anything 3 and chained into a common relative frame. Learned depth is defined only up to scale, so each segment is pinned to metres by the phone's co-registered LiDAR depth; segments without enough metric support are rejected rather than forced. Future poses are exported as metric trajectory targets paired with visual goals.
Every part is off-the-shelf and needs no machining or custom electronics. Prices are what each listing showed on , before tax and shipping.
| Component | Item | Cost |
|---|---|---|
| Rollator walker | Helavo all-terrain rollator walker, 12″ pneumatic tires, foldable, 300 lb capacity | $229.90 |
| Phone mount | Lamicall 1s-lock handlebar phone mount, camera-friendly, 4.7–7″ phones | $21.99 |
| Smartphone (RGB + LiDAR) | Consumer smartphone with a rear LiDAR scanner; an iPhone 16 Pro was used throughout for consistency | Operator-supplied |
| Total | Excluding the smartphone | $251.89 |
Assembly takes about 30 minutes. No 3D-printed parts, no custom brackets, no soldering.
The complete build, time-lapsed; the overlay counts elapsed real time, 30:09 end to end. Faces and the background are blurred to preserve the anonymity of this submission — the hardware and the assembly steps themselves are unmodified.
An egocentric, rollator-constrained navigation corpus recorded in public pedestrian space: crowded sidewalks, building entrances, transit interfaces, wet pavement, and construction-constrained walkways. It keeps the parts of navigation that are usually filtered out — waiting at crossings, yielding to pedestrians, stop–resume transitions — because those are supervision too.
| Distance | 37.2 km |
| Recording time | 10.8 h |
| Sessions | 87 |
| Operators | 6 |
| Locations | Three cities in the United States |
| Collection period | Intermittently, April–August 2026; collection is ongoing |
| Near-stationary windows | 8.14% |
| Model-ready export | 1,087 trajectories, 117,512 frames at ≈4 Hz |
| Surface types | Concrete sidewalks, brick and tactile paving, asphalt paths and roads, gravel, cobblestone, boardwalks, dirt, grass, sand, carpet, tile, terrazzo, with examples of wet pavement |
| Also captured | Transit by train and elevator, yielding to pedestrians, waiting at crosswalks |
The initial collection target was 10 h, set to evaluate the pipeline and its utility for navigation learning. Two UNI systems are now in use for parallel collection, so the released corpus is expected to extend beyond this snapshot. Every result on this page uses the fixed snapshot above. A window counts as near-stationary when its cumulative 2-D path length over five consecutive frame-to-frame steps is below 0.25 m; on that same criterion EgoWalk is 3.10% and SCAND 0.25%.
Coming Soon