# Camera Debug Summary

## Executive Finding

自采 RoboTwin HDF5 里已经保存了 camera 参数；WorldArena val/test input HDF5 没保存 camera 参数。当前 action map overlay 偏移的第一层 bug 是 converter 之前没有读取自采 HDF5 的 camera，而是无条件写 `camera_fallback`。这个已经在 `convert_robotwin_to_manifest.py` 修复。

## HDF5 Field Inspection

### Self-collected RoboTwin HDF5

Root: `/root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0`
Files scanned: 50

- `observation/head_camera/intrinsic_cv`: 50/50
- `observation/head_camera/extrinsic_cv`: 50/50
- `observation/head_camera/cam2world_gl`: 50/50

Detected cameras include `head_camera` and `front_camera`. The relevant fields are per-frame arrays:

- `observation/head_camera/intrinsic_cv`: (124, 3, 3): 1, (128, 3, 3): 1, (120, 3, 3): 2, (119, 3, 3): 1, (79, 3, 3): 5, (101, 3, 3): 1
- `observation/head_camera/extrinsic_cv`: (124, 3, 4): 1, (128, 3, 4): 1, (120, 3, 4): 2, (119, 3, 4): 1, (79, 3, 4): 5, (101, 3, 4): 1
- `observation/head_camera/cam2world_gl`: (124, 4, 4): 1, (128, 4, 4): 1, (120, 4, 4): 2, (119, 4, 4): 1, (79, 4, 4): 5, (101, 4, 4): 1

### WorldArena Val HDF5

Root: `/root/autodl-tmp/worldarena_testset/val_dataset/data/fixed_scene_task`
Files scanned: 50

- `observation/head_camera/intrinsic_cv`: 0/50
- `observation/head_camera/extrinsic_cv`: 0/50
- `observation/head_camera/cam2world_gl`: 0/50

WorldArena input HDF5 does not expose camera matrices in the checked val data. It mainly exposes action/endpose/gripper style fields.

## RoboTwin Collection Code

RoboTwin already writes runtime camera fields:

- `/root/autodl-tmp/RoboTwin/envs/camera/camera.py`: `Camera.get_config()` reads `get_intrinsic_matrix()`, `get_extrinsic_matrix()`, and `get_model_matrix()`.
- `/root/autodl-tmp/RoboTwin/envs/_base_task.py`: `Base_Task.get_obs()` stores `self.cameras.get_config()` under `observation`; `_take_picture()` writes frame PKLs; `merge_pkl_to_hdf5_video()` merges them.
- `/root/autodl-tmp/RoboTwin/envs/utils/pkl2hdf5.py`: `create_hdf5_from_dict()` recursively writes nested fields to HDF5.

Detailed report: `/root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0/camera_inspection/camera_save_code_report.md`.

## Converter Fix

`/root/autodl-tmp/worldarena_data_factory/convert_robotwin_to_manifest.py` now reads camera fields in this order:

1. `observation/head_camera/intrinsic_cv` + `observation/head_camera/extrinsic_cv`
2. `observation/head_camera/cam2world_gl` if `extrinsic_cv` is unavailable
3. equivalent `front_camera` fields
4. fallback only if no camera fields exist

For ABot/VACE JSON compatibility, `extrinsic_cv` is treated as OpenCV world-to-camera and inverted before writing `camera_extrinsic.json`, because the ABot action-map code expects camera-to-world and then internally computes world-to-camera.

New manifest fields:

- `camera_source`
- `camera_name`
- `camera_convention`
- `camera_info_path`

New quality flag when successful: `camera_hdf5_verified`.

## Validation Counts

Current full manifest camera status before full reconvert:

```json
{
  "manifest_missing": "/root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0/manifests/episode_manifest.parquet"
}
```

Small reconvert validation on `grab_roller/wa_clean_fixed__aloha_agilex__job_00128`:

```json
{
  "hdf5_verified": 5
}
```

The small reconvert produced 5/5 `camera_source=hdf5_verified`.

## Overlay Compare

Generated outputs:

- HDF5 camera overlay: `/root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0/camera_debug_job00128_hdf5cam`
- GL convention test overlay: `/root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0/camera_debug_job00128_glcam`
- fallback vs HDF5 compare: `/root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0/camera_compare_job00128`
- compare report: `/root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0/camera_compare_job00128/camera_compare_report.md`

Result: HDF5 camera is now loaded, but several action maps are still empty/out-of-frame or near image corners. Using `cam2world_gl` directly did not solve it. That suggests the remaining offset is probably not just converter fallback; likely causes are:

- ee16/endpose coordinate frame is not the exact frame ABot/VACE projection assumes;
- EE pose origin is wrist/tool/end-effector center, not gripper fingertip/contact point;
- left/right arm or OpenCV/OpenGL convention still needs a deeper coordinate-frame calibration test;
- action map may intentionally encode projected EE trajectory rather than pixel-perfect gripper silhouette, so exact fingertip overlap is not required.

## Need To Recollect?

For camera parameters: no, not solely for camera. Existing self-collected HDF5 already contains camera matrices and can be repaired by rerunning converter.

For training-quality action maps: do not trust full A2V until the ee16 coordinate-frame/projection sanity check is resolved. SFT video-only data is unaffected by camera projection. A2V with action maps should wait for a 10-30 sample manual action-map check after reconversion.

## Next Commands

Reconvert an existing raw-data root after collection finishes:

```bash
/root/autodl-tmp/conda/envs/fantasyworld/bin/python /root/autodl-tmp/worldarena_data_factory/convert_robotwin_to_manifest.py \
  --out /root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0_reconverted \
  --robotwin-root /root/autodl-tmp/RoboTwin \
  --jobs-csv /root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0/manifests/robotwin_collection_jobs.csv
```

Then run action-map visualization on 10 samples:

```bash
/root/autodl-tmp/conda/envs/fantasyworld/bin/python /root/autodl-tmp/worldarena_data_factory/visualize_abot_action_maps.py \
  --manifest /root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0_reconverted/manifests/episode_manifest.parquet \
  --abot-root /root/autodl-tmp/ABot-PhysWorld \
  --out /root/autodl-tmp/worldarena_data_factory_v0_clean_rt256_gpu0_reconverted/action_map_visualization_test \
  --num-samples 10 \
  --sample-by task_family \
  --representation ee16 \
  --alpha 0.55 \
  --seed 42
```