sim_data / dresser /README.md
Ravenh97's picture
dresser README
60f614e verified
|
Raw
History Blame Contribute Delete
1.65 kB
# Dresser sim data (open top drawer)
Task: open the top drawer of `ithor_minimal_dresser` (asset `Drawer_b7714dd0`)
with the DROID arm + planner policy. One chunk = one slurm job = one 1-GPU shard.
## Layout
```
dresser/sim/chunk_NNN/
trajectories_batch_1_of_1_cam_rand.h5
episode_NNNNNNNN_exo_camera_1_robot_rgb_batch_1_of_1_cam_rand.mp4
episode_NNNNNNNN_wrist_camera_robot_rgb_batch_1_of_1_cam_rand.mp4
episode_NNNNNNNN_*_depth.npz
episode_NNNNNNNN_*_segmentation.npz
chunk_log.txt
```
Each chunk has a `chunk_log.txt` recording its setup (seed, node, args).
## Sim arguments
From `molmospaces/scripts/datagen/data_gen_dresser.sh`:
- scene_dataset: `ithor_minimal_dresser`
- target: drawer, `object_instance_index=0`, `joint_index=0` (top drawer)
- camera_extrinsics: `picked_camera_extrinsics.json` + pos noise 0.05, rot noise 5°
- base_pose_sampling_radius_range: 0.8–1.1 m
- base_pose_sampling_theta_range_rad: 1.6–2.8 (≈92°–160°, front-left wedge)
- robot_placement_rotation_range_rad: 0.5
- robot_base_z: 0.86 (with `z_jitter -0.2 -0.1` → effective 0.66–0.76 m)
- robot_init_qpos_noise_range: 0.4 on all 7 joints
## Pipeline
- Each chunk runs `run_pipeline_record_state.py` with `samples_per_house=12`,
`batch_num=1`, `total_batches=1`, `filter_for_successful_trajectories`.
- After sim exits, `check_chunk.py` verifies `T == exo_mp4_frames == wrist_mp4_frames`
for every traj. Chunk is pushed only if check passes.
- Push uses `huggingface_hub.HfApi.upload_folder` (one commit per chunk).
## Naming
- `chunk_NNN` where `NNN = SLURM_ARRAY_TASK_ID` (1-padded to 3 digits).
- `seed = 20000 + chunk_id`.