File size: 7,302 Bytes
180c67c 7684a18 180c67c 7684a18 53f3821 53f7fe6 7684a18 180c67c aa62c5f 180c67c 7684a18 180c67c 7684a18 180c67c 7684a18 180c67c 7684a18 180c67c 7684a18 180c67c 7684a18 180c67c 7684a18 180c67c 53f3821 7684a18 180c67c 7684a18 180c67c 7684a18 180c67c aa62c5f 180c67c aa62c5f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 | # Data quality
A small fraction of frames contain known sensor artifacts. The repo ships [`../bad_frames.json`](../bad_frames.json) as a skip-list so downstream code can avoid sampling on top of these intervals — important since this dataset is intended for **short-window dynamics / world-model learning**, where a glitch landing inside a training window can dominate the loss for that step.
> **All frame indices in `bad_frames.json` are in TRIMMED `.pt` coordinates.** The original H5 recordings have not been edited; only the published `.pt` files have had their OT-uninitialized prefixes cut (see [`caveats.md`](caveats.md) §OT track loss). The per-file trim offset is stored in `_contact_meta.trim_offset` inside each `.pt`.
## Headline numbers
| | |
|---|---:|
| Total synchronized frames (post-trim) | 190,231 (105.7 min @ 30 Hz) |
| Recording files | 27 |
| Frames flagged in `bad_frames.json` | **1,768 (0.929 %)** |
| Recording files with ≥ 1 flagged frame | 11 |
## Failure modes — quantitative breakdown
| # | Mode | Frames | % of dataset | Files | Symptom | Example |
|---|---|---:|---:|---:|---|---|
| 1 | **GelSight LED flicker** | 56 | **0.029 %** | 5 / 27 | 1–2 frames of uniform pink/magenta wash across the gel surface; adjacent frames normal. | [`intensity_spike_overview.png`](../figures/dataset_figures/intensity_spike_samples/intensity_spike_overview.png) |
| 2 | **OptiTrack pose teleport** | 56 | **0.029 %** | 3 / 27 | Translation velocity > 5 m/s **or** angular velocity > 15 rad/s in a single OT sample (solver flip — physically impossible for human-hand motion). | [`pose_teleport_samples/`](../figures/dataset_figures/pose_teleport_samples/) |
| 3 | **OptiTrack track loss** | 1,680 | **0.883 %** | 6 / 27 | Sensor pose held at a stale value for ≥ 0.25 s; cross-modal motion check confirms hand was actually moving. Cause: marker briefly left the mocap volume / camera FOV. | [`freeze_diagnose/`](../figures/dataset_figures/freeze_diagnose/) |
**Union of modes 1+2+3: 1,768 / 190,231 = 0.929 % of post-trim frames flagged.** Mode 3 is by far the largest — but it's already orders of magnitude smaller than what the *raw* recordings looked like (15 % of frames before the OT-uninitialized prefixes were trimmed). The remaining 0.88 % is genuine mid-episode mocap dropout that no amount of post-processing can recover.
## How to use the data — three recipes
### A. Easiest: the shipped example dataloader
`ReactWindowDataset` does the right thing out of the box. Just turn `skip_bad_frames=True`:
```python
from examples.react_window_dataset import ReactWindowDataset
from torch.utils.data import DataLoader
ds = ReactWindowDataset(
data_root = "processed/mode1_v1/motherboard",
bad_frames_path = "bad_frames.json",
tasks_json_path = "tasks.json",
window_length = 16,
skip_bad_frames = True, # ← drops windows touching ANY of modes 1/2/3
respect_active_sensors = True,
)
```
Any window whose [t_start, t_end] range overlaps an `intensity_spikes` /
`pose_teleports_{L,R}` / `ot_loss_{L,R}` interval is silently dropped at
window-enumeration time, so your DataLoader never sees them.
### B. Loading a single `.pt` directly — DIY skip-list
If you're rolling your own sampler or scanning a single file:
```python
import json, torch
ep_key = "2026-05-11/episode_017"
ep = torch.load(f"processed/mode1_v1/motherboard/{ep_key}.pt", weights_only=False)
T = ep["view"].shape[0]
trim_offset = ep["_contact_meta"].get("trim_offset", 0) # already applied to ep
bad = json.load(open("bad_frames.json"))["episodes"][ep_key]
# Build a per-frame boolean mask (True = drop)
import numpy as np
mask = np.zeros(T, dtype=bool)
for s_, e_ in (bad["intensity_spikes"]
+ bad["pose_teleports_L"] + bad["pose_teleports_R"]
+ bad["ot_loss_L"] + bad["ot_loss_R"]):
mask[s_:e_ + 1] = True
print(f"{mask.sum()}/{T} flagged ({100 * mask.mean():.2f} %)")
# Then sample only clean windows:
def clean_window_start(t_start, win_len):
return not mask[t_start:t_start + win_len].any()
```
### C. Only care about action labels (pose)? Skip just ot_loss
If your model treats vision + tactile as observations and sensor pose as the *action* label (UMI-style imitation learning), the LED flicker and translation teleports are observation-side noise that you can usually tolerate (single frames or short < 0.3 s spikes). The OT track losses are the one mode you *must* exclude, because the recorded "action" is stale:
```python
# Just the action-label cuts
def clean_pose_window(t_start, win_len, bad_ep):
for s_, e_ in (bad_ep["ot_loss_L"] + bad_ep["ot_loss_R"]
+ bad_ep["pose_teleports_L"] + bad_ep["pose_teleports_R"]):
if s_ <= t_start + win_len - 1 and e_ >= t_start:
return False
return True
```
### What NOT to do
- Don't ignore `bad_frames.json` — for `ot_loss` intervals the *recorded pose looks plausible* (it's just a stale held value), but the actual sensor was moving. A model trained on those will learn that contact-rich tactile signals are uncorrelated with motion.
- Don't try to skip individual *frames* and stitch the rest — windows must be contiguous over time. Drop the **whole window**, not just the bad frames in it.
## Inspection figures
**Mode 1 — GelSight LED flicker** (overview across all affected files):

Per-file close-ups: [`figures/dataset_figures/intensity_spike_samples/`](../figures/dataset_figures/intensity_spike_samples/).
**Mode 2 — OptiTrack pose teleport.** GIFs (10 s playback at 2× speed) for all affected files: [`pose_teleport_samples/`](../figures/dataset_figures/pose_teleport_samples/).
**Mode 3 — OptiTrack track loss.** 15 sample clips (recording-viewer layout, 30 fps real time, red ring on the frozen-sensor projected dot): [`freeze_diagnose/ot_loss/`](../figures/dataset_figures/freeze_diagnose/ot_loss/).
## `bad_frames.json` schema
Frame indices are inclusive on both ends and pre-padded by `buffer_frames` (3) on each side so context windows don't bleed into the glitch:
```json
{
"tau_intensity": 30.0,
"tau_velocity_mps": 5.0,
"tau_angular_rad_per_s": 15.0,
"tau_opt_gap_s": 0.10,
"freeze_threshold_s": 0.25,
"buffer_frames": 3,
"summary": {
"n_episodes": 27,
"total_frames": 190231,
"total_bad_frames": 1768,
"bad_fraction_overall": 0.009294,
"n_episodes_with_bad_frames": 11
},
"episodes": {
"2026-05-11/episode_003": {
"n_frames": 10032,
"duration_s": 334.4,
"intensity_spikes": [[260, 268], "..."],
"pose_teleports_L": [],
"pose_teleports_R": [],
"ot_loss_L": [],
"ot_loss_R": [],
"total_bad_frames": 19,
"bad_fraction": 0.0019
}
}
}
```
A per-mode aggregate is also available: [`figures/dataset_figures/data_quality_breakdown.json`](../figures/dataset_figures/data_quality_breakdown.json).
## Full report (per-file CSV)
Every file's individual stats (frames, duration, contact %, max intensity left/right, drift, max pose velocity, etc.) are in [`figures/dataset_figures/data_quality_report.csv`](../figures/dataset_figures/data_quality_report.csv).
|