# Rabbit-RobotNav — room survey and heading evidence

Updated 9 October 2026. This is a presentation update, not a new robot experiment. The page preserves earlier negative results and adds completed, traceable room-survey and heading-investigation records.

## Frozen research source

Repository: [history-aware-mp3d-vln / real-robot-integration](https://github.com/lijinghai/history-aware-mp3d-vln/tree/96d146e0d9822cdd71e601b2b376057cedd8f147), commit `96d146e0d9822cdd71e601b2b376057cedd8f147`. The initial page update used its parent `18957ee8ec62470ac2d57dbe72c6c7ab0d437dc9` for the completed survey, geometry audit, actual isolated EKF A/B, and driver covariance audit. The IMU-chain task completed during website delivery; its committed stationary findings were then checked against the source report and remote SHA and added briefly. No ongoing or unverified turning work is reported as complete. The research worktree, robot configuration, and running experiments were not modified.

Evidence directories below are relative to `docs/evidence/chassis/wheeltec-mecanum/` in that pinned snapshot.

## Selected images

| Website asset | Origin / role | SHA-256 |
| --- | --- | --- |
| `assets/research-framework-2026-10-09.png` | User-supplied original framework, 1280 × 720; conceptual design, not a runtime screenshot | `f5b376c52bba3f26912a914db3e84fa2ccf86dffca4c0b5b56f93579a96db693` |
| `assets/room-survey-2026-10-09.png` | `room-map-v1-live-survey-2026-10-09/room_semantic_map.png`, 1960 × 1260; recorded occupancy grid, estimated trajectory and original RGB captures | `e96edf4b7f0ff46a565d9d60661871a11d721201a2eb3276fcc9684b9aff0aee` |
| `assets/ekf-heading-ab-2026-10-09.png` | `ekf-isolated-heading-ab-v1-2026-10-09/ekf_yaw_comparison.png`, 1920 × 1120; actual isolated EKF output compared with recorded inputs | `228e11eed44c99fa691f0334730eb5cb445454af0e35547c218a2c423904be4d` |

All three files are byte-for-byte copies, with source and destination SHA-256 checked. No cropping, recoloring, synthetic scene reconstruction, or regenerated measurements were used. The supplied framework is kept as supplied, including its module names and blue/purple colors; the surrounding website retains its orange/navy theme. Its detector, multimodal reasoning, trusted-memory and recovery blocks describe the intended design, not an assertion that every module has been integrated and validated.

## What changed since 8 October

### Manual room survey: captured, not accepted

The earlier stationary preflight did not produce an occupancy grid. The subsequent **human-controlled** survey recorded approximately 229 seconds, 2,747 laser scans, 4,580 raw odometry messages, 2,281 fused odometry messages, and 225 RGB keyframes (224 with camera intrinsics). A 427 × 584 occupancy grid at 0.05 m resolution was obtained. This replaces the outdated statement that the survey is still pending. [Capture protocol](https://github.com/lijinghai/history-aware-mp3d-vln/tree/18957ee8ec62470ac2d57dbe72c6c7ab0d437dc9/docs/evidence/chassis/wheeltec-mecanum/room-map-v1-live-survey-2026-10-09).

The status remains `ROOM_GEOMETRY_CAPTURED_PENDING_REVIEW` / `ROOM_MAP_V1=PENDING_REVIEW`. Radial streaks, sparse wall structure and inconsistent return estimates require review. The source image's filename `room_semantic_map.png` does **not** establish semantic mapping: the detector was `NOT_RUN`, no physical instance was verified, and trusted metric landmarks remain zero. Its “measured odom path” label refers to integrated odometry estimates, not independently measured physical distance. The survey is separate from the 1,149-state archived history replay.

### Heading investigation: failed candidate, supported source attribution

The geometry audit found that fused yaw and angular velocity remained zero while raw odometry recorded turns. Map artifacts do not have an independently established single cause. [Geometry audit](https://github.com/lijinghai/history-aware-mp3d-vln/tree/18957ee8ec62470ac2d57dbe72c6c7ab0d437dc9/docs/evidence/chassis/wheeltec-mecanum/room-map-v1-geometry-audit-2026-10-09).

The subsequent A/B used the actual ROS 2 `robot_localization` EKF in network-isolated runs with the same 4,580 recorded odometry messages. The only research variable was `odom0_config[11]`, enabling angular-z velocity fusion. Candidate yaw changed by −0.077352249 rad, compared with −7.121530533 rad in recorded raw odometry and zero in the isolated baseline. Rounded values on the page are −0.077 and −7.122 rad. The status is `EKF_ISOLATED_AB_CANDIDATE_FAILED`, not “heading fixed.” These fresh-start runs are not a full recreation of historical filter initialization; the survey bag lacks synchronized IMU data. Raw odometry is not orientation ground truth. No candidate was deployed to the robot. [A/B protocol](https://github.com/lijinghai/history-aware-mp3d-vln/tree/18957ee8ec62470ac2d57dbe72c6c7ab0d437dc9/docs/evidence/chassis/wheeltec-mecanum/ekf-isolated-heading-ab-v1-2026-10-09).

The read-only covariance audit matched the driver source, executable constants, and all 4,580 recorded messages: when any of the three velocity components is nonzero, `twist.covariance[35]` is 1,000 (rad/s)²; when all are exactly zero it is 1e−9. This supports strong down-weighting of angular velocity in the tested candidate, not a calibrated replacement value or a validated fix. Pose-yaw variance and angular-velocity variance have different units and are not interchangeable. The source snapshot was inspected at `Publish_Odom` and the four covariance arrays. [Covariance audit and source](https://github.com/lijinghai/history-aware-mp3d-vln/tree/18957ee8ec62470ac2d57dbe72c6c7ab0d437dc9/docs/evidence/chassis/wheeltec-mecanum/wheeltec-angular-covariance-audit-2026-10-09).

### Stationary IMU audit: messages available, turning unverified

The separate stationary audit at `96d146e` recorded 500 raw IMU messages at approximately 20 Hz. The configured EKF input `/imu/data_filtered` had no publisher. `IMU_CHAIN_EVIDENCE_PARTIAL` means messages were received and static descriptive screening was completed, not a passed sensor calibration: turning response and independent measurement reliability remain unverified, and the next fusion candidate is `NONE`. The original quaternion norm was approximately 0.998313; driver covariance constants are not calibrated precision. This stationary session must not be combined with the earlier room-survey turns, which lack synchronized IMU. The next controlled-turn experiment requires separate field authorization and an independent angle reference. To keep the page concise, its original plots remain linked in the [source report](https://github.com/lijinghai/history-aware-mp3d-vln/tree/96d146e0d9822cdd71e601b2b376057cedd8f147/docs/evidence/chassis/wheeltec-mecanum/wheeltec-real-imu-chain-v1-2026-10-09) rather than adding another large image.

## Preserved evidence boundaries

Earlier MP3D, CMA, chair-identity and real-history findings retain their own denominators, dates and limitations; see [8 October provenance](evidence-2026-10-08.md). No observed results have been removed because they were negative. The 8 October live diagnostic initialization failure is dated explicitly rather than described as the latest robot run. Reliable geometry, trusted landmarks, benchmark improvement and autonomous recovery are still unverified. The project remains an unpublished master's thesis project.

## Editable research outline

The [Mermaid mind map](research-mindmap.mmd) was updated using the Axton `mermaid-visualizer` workflow and rendered with the existing orange/navy configuration to a [local SVG](assets/research-mindmap-2026-10-09.svg). It connects completed capture, failed fusion, the IMU validation gate, and the research question. It is an explanatory outline, not additional experimental evidence. The browser loads the rendered SVG without an external Mermaid runtime.