# Dreame 06-03 SpatialLM voxel comparison

This dataset contains Rerun recordings and SpatialLM text outputs for the same
colored Open3D TSDF point cloud reconstructed from the Dreame 06-03 active
stereo sequence. It compares SpatialLM point-cloud cleanup voxel sizes of
`0.075 m` and `0.025 m`.

## Files

| Path | Description |
|---|---|
| `active_stereo_tsdf_point_cloud.rrd` | Existing historical `0.075 m` Rerun recording. |
| `runs/voxel_0.075/` | Historical layout, raw generation, and diagnostics. |
| `runs/voxel_0.025/` | New layout, raw generation, diagnostics, and standalone Rerun recording. |
| `comparison/active_stereo_tsdf_voxel_0.075_vs_0.025.rrd` | Three-view comparison: `0.075`, `0.025`, and overlay. |
| `comparison/comparison_report.json` | Input hashes, parameters, entity counts, runtime summaries, and coordinate audit. |

Open the comparison with Rerun 0.21 or a compatible viewer:

```bash
rerun comparison/active_stereo_tsdf_voxel_0.075_vs_0.025.rrd
```

The input point cloud is logged once in its original colors. The `0.075`
layout is blue, the `0.025` layout is orange, and both can be toggled
independently. The recording explicitly uses a right-handed Z-up coordinate
frame and an XY grid.

## Input provenance

- Source artifact: `active_stereo_tsdf_point_cloud.ply`
- SHA-256: `cfcfd27fae0174b4b69616a18c7c50e8fd4f19a4d8f0679624c15c51c207dc3e`
- Points: 175,025, with RGB colors and normals
- Unit: meters
- Coordinate transform applied before inference: identity
- Surface-normal Manhattan yaw audit: `-0.4434 degrees`; no additional rotation was applied

The source PLY and model weights are not distributed in this dataset.

## Inference settings

Both results use `SpatialLM1.1-Qwen-0.5B`, `detect_type=all`, seed `42`, MPS
for the model and Sonata backbone, mtlgemm sparse convolution, SDPA attention,
FP16 LLM weights, block synchronization, 16-block convolution chunking, and a
maximum of 4096 generated tokens. The SpatialLM model grid is `0.025 m` in
both runs.

| Metric | `0.075 m` cleanup | `0.025 m` cleanup |
|---|---:|---:|
| SpatialLM Mac commit | `8913c44` | `ba67311` |
| Points after cleanup | 29,460 | 119,767 |
| Tokens after grid sampling | 29,166 | 101,697 |
| Generated tokens | 896 | 946 |
| Walls | 19 | 14 |
| Doors | 5 | 7 |
| O-BBoxes | 5 | 10 |
| Total parsed entities | 29 | 31 |

All five Sonata stages completed on MPS with finite outputs. Irregular sparse
topology operations, including neighbor-map construction, serialization, and
large batch-offset reductions, use explicit and recorded CPU helpers for
watchdog safety and MPS correctness. `PYTORCH_ENABLE_MPS_FALLBACK` was not
enabled.

## Comparison limitation

This is a historical visual comparison, not a strict single-variable
ablation. The `0.075 m` result is intentionally reused from the existing
recording, while the `0.025 m` result uses the later Mac backend commit that
fixes large MPS point-batch offsets. Differences may therefore reflect both
the cleanup voxel size and the backend revision. No claim is made that either
layout is geometrically correct.