Ravenh97 commited on
Commit
f9d1e6d
·
verified ·
1 Parent(s): 3508efb

fridge_m_v2: initial SETUP.md

Browse files
Files changed (1) hide show
  1. fridge_m_v2/SETUP.md +102 -0
fridge_m_v2/SETUP.md ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # fridge_m_v2 sim datagen — setup record
2
+
3
+ Generated 2026-05-14 by an automated agent on viscam6 (4× RTX A5000), with
4
+ helper nodes viscam7 (4× A5000) and viscam3 (4× RTX 3090). Sim-only run;
5
+ no concurrent gen on this dataset. Push target:
6
+ `Ravenh97/sim_data:fridge_m_v2/`.
7
+
8
+ ## Motivation
9
+
10
+ Regenerate fridge_m sim data after the user updated:
11
+ - The scene XML (`assets/scenes/ithor_minimal_fridge_m/FloorPlan1_physics.xml`)
12
+ — modified 2026-05-14 ~20:56 PDT.
13
+ - The robot init params and placement ranges in
14
+ `scripts/datagen/data_gen_fridge_m_2.sh` — modified 2026-05-14 ~21:02 PDT.
15
+
16
+ ## Pipeline
17
+
18
+ Same chunked sim pipeline that worked for fridge_m / fridge_7 / 3drawers:
19
+ - Each chunk runs 4 sim shards on 4 GPUs writing to `sim_chunks/chunk_NNN/`.
20
+ - After all 4 shards exit (h5 locks released), `check_mismatch.py` asserts
21
+ `T == exo_mp4_frames == wrist_mp4_frames` for every traj.
22
+ - NO gen step (sim-only request).
23
+
24
+ Frozen scripts are at `/scr/ravenh/fridge_m_v2/scripts/` on each node so
25
+ mid-run edits to the source `data_gen_fridge_m_2.sh` cannot silently
26
+ change the run. The frozen `data_gen_v2.sh` only differs from the source
27
+ by absolutizing the `PIPELINE` path (snapshot lives under `/scr` so the
28
+ SCRIPT_DIR-relative path in the original would not resolve).
29
+
30
+ ## v2 parameters (from frozen `data_gen_v2.sh`)
31
+
32
+ ```
33
+ --scene_dataset ithor_minimal_fridge_m
34
+ --object_instance_index 1
35
+ --joint_index 1
36
+ --base_pose_sampling_radius_range 0.8 1.2
37
+ --base_pose_sampling_theta_range_rad -0.5 1.0
38
+ --robot_placement_rotation_range_rad 0.5
39
+ --robot_placement_exclusion_threshold 0.0
40
+ --robot_base_z 0.86
41
+ --robot_base_z_jitter -0.1 0.1
42
+ --robot_init_qpos_noise_range 0.4 0.4 0.4 0.4 0.4 0.4 0.4
43
+ --camera_extrinsics scripts/datagen/picked_camera_extrinsics.json
44
+ --camera_extrinsics_pos_noise 0.05
45
+ --camera_extrinsics_rot_noise_deg 5.0
46
+ --output_suffix _cam_rand
47
+ --record_depth, --record_segmentation, --record_robot_masked_rgb
48
+ --filter_for_successful_trajectories
49
+ ```
50
+
51
+ Pipeline: `molmospaces/scripts/datagen/run_pipeline_record_state.py`
52
+ (records full mocap/qpos state per step; same wrapper as fridge_m_0.7_pd).
53
+
54
+ ## Diff vs fridge_m (v1)
55
+
56
+ | param | v1 (data_gen_fridge_m.sh) | v2 (data_gen_fridge_m_2.sh) |
57
+ | ------------------------------------- | ------------------------- | --------------------------- |
58
+ | base_pose_sampling_radius_range | 0.75 1.1 | 0.8 1.2 |
59
+ | base_pose_sampling_theta_range_rad | -0.5 0.9 | -0.5 1.0 |
60
+ | robot_placement_rotation_range_rad | 0.2 | 0.5 |
61
+ | robot_base_z_jitter | 0.0 0.1 | -0.1 0.1 |
62
+ | robot_init_qpos_noise_range | (unset) | 0.4 × 7 |
63
+ | pipeline | run_pipeline_no_clip.py | run_pipeline_record_state.py|
64
+
65
+ Scene XML was also modified in place — fingerprints recorded below.
66
+
67
+ ## Fingerprints
68
+
69
+ - `assets/scenes/ithor_minimal_fridge_m/FloorPlan1_physics.xml` md5 `83a5b0c9d5a28208343e70c2b1151af6`
70
+ - `assets/scenes/ithor_minimal_fridge_m/FloorPlan1_physics.json` md5 `1670a043ef6ddfd3c872d9fa221a7bbe`
71
+ - `scripts/datagen/data_gen_fridge_m_2.sh` md5 `86be8ec056ebbcb8668c2f80a4ed8eeb`
72
+ - frozen `data_gen_v2.sh` (absolute PIPELINE path) md5 `cf816d56054141e9ea8e7b51226623cd`
73
+ - molmospaces git HEAD `8003c7009d28746621f969ea44e96475f851d2cd`
74
+
75
+ ## Node assignments
76
+
77
+ | node | GPUs | chunk range | SEED_BASE | STOP_AT_TRAJS | /scr free at launch |
78
+ | -------- | ----------------- | ----------- | --------- | ------------- | ------------------- |
79
+ | viscam6 | 4× A5000 (0-3) | 001-015 | 20000 | 370 | 5.1 TB |
80
+ | viscam7 | 4× A5000 (0-3) | 101-115 | 30000 | 370 | 3.3 TB |
81
+ | viscam3 | 4× RTX 3090 (0-3) | 201-210 | 40000 | 260 | 248 GB |
82
+
83
+ Expected per-chunk yield (from fridge_m_chunked memory): ~45-50 trajs.
84
+ Target ~1000 total.
85
+
86
+ ## HF layout
87
+
88
+ ```
89
+ Ravenh97/sim_data:fridge_m_v2/
90
+ SETUP.md <- this file
91
+ sim_chunks/
92
+ chunk_001/ ... <- viscam6
93
+ chunk_002/ ...
94
+ ...
95
+ chunk_101/ ... <- viscam7
96
+ ...
97
+ chunk_201/ ... <- viscam3
98
+ ...
99
+ ```
100
+
101
+ push_watchdog.py uploads (1 commit per push) every 50 finalized trajs;
102
+ skips h5 files until the final flush (then includes them).