React / docs /figures.md
yxma's picture
Audit batch 1: post-trim numbers in docs/figures.md (#3); tasks.json 2026-05-11 note + trim_offsets lookup (#5, #8); bad_frames.json trim_offsets header (#8); dataloader default which_sensors='both' (#10); new freeze_intervals.json with per-event diagnostic signals (#12)
70ded71 verified
# Statistics and analysis
A gallery of dataset-level figures. All sources live under [`figures/dataset_figures/`](../figures/dataset_figures/).
## F1 — Episode length distribution
![Episode length](../figures/dataset_figures/F1_episode_length_histogram.png)
Across 27 episodes, total duration **106 min**. Episode lengths span 25 s to ~19 min (median ~4 min) — about **8× longer than BridgeData V2's typical 30 s demo**.
## F2 — Contact event durations
![Contact event durations](../figures/dataset_figures/F2_contact_event_duration_histogram.png)
Across the 27 post-trim recording files, ~**146,440** frames (77.0%) are in tactile contact. The typical contact event lasts ≈ **0.7 s** (median); tail out to 33 s.
## F3 — Where on the gel does contact land?
![Contact heatmap](../figures/dataset_figures/F3_contact_heatmap.png)
Both sensors show contact concentrated in the central ~50 % of the gel surface, dropping toward the edges. The left gel has discrete bright spots from repeated contacts with specific features.
## F4 — Tactile signal is real and varied
![Tactile montage](../figures/dataset_figures/F4_tactile_pattern_montage.png)
16 random contact frames sampled across the dataset — discrete pins, edges, smooth surfaces, multi-object contact.
## F5 — Bimanual workspace
![Pose trajectory](../figures/dataset_figures/F5_pose_trajectory_3d.png)
Multi-view projection of the longest episode (`2026-05-11/episode_017`, 19 min). Left (blue) and right (orange) sensors operate over a ~30 × 40 × 30 cm workspace.
## F6 — Tactile is independent of motion
![Cross-modal correlation](../figures/dataset_figures/F6_cross_modal_correlation.png)
Sensor velocity vs tactile intensity is **essentially uncorrelated** (r ≈ +0.04 / −0.05). Tactile carries information that is **not** explained by pose+velocity — a direct argument for the value of including tactile in policy / world-model training.
## F7 — Comparison with other manipulation datasets
![Comparison table](../figures/dataset_figures/F7_comparison_table.png)
React is the only entry combining tactile + RGB-D + motion capture + bimanual.
## F8 — Per-episode summary
![Per-episode summary](../figures/dataset_figures/F8_per_episode_summary.png)
Full table for all 27 episodes. CSV: [`figures/dataset_figures/F8_per_episode_summary.csv`](../figures/dataset_figures/F8_per_episode_summary.csv).
## Tactile-intensity timelines
![Full dataset intensity timeline](../figures/contact_intensity_full.png)
106-minute audio-waveform-style view across the entire dataset; left intensity rendered above the x-axis (orange), right intensity mirrored below (blue). A 20-minute zoom is in [`figures/contact_intensity_20min.png`](../figures/contact_intensity_20min.png).
> **Note**: The hero figures above (`contact_intensity_*.png`) and `F1`–`F8` are pre-trim renders; their x-axis durations are slightly overstated for `2026-05-11/episode_{005,012,017}`. They will be regenerated in a follow-up commit. Headline numbers (frames, duration, bad-frame %) in `README.md` and `docs/quality.md` are post-trim and authoritative.