# Trajectory deduplication (`dedup/`)

Near-duplicate removal over the 94,030-episode DROID LeRobot export
(94,006 episodes with valid kinematic signals were assessed; 24 episodes with
degenerate signals are kept by default). A duplicate is a **near-identical
re-recording of the same task**; different tasks — including paraphrases
resolved to different tasks, object swaps, or the same task done in a
different place — are treated as diversity and never merged.

## Method

1. **Merge task phrasings** — instructions are canonicalised (synonym folding
   for verbs/objects, prepositions to `<p>`, articles dropped, order and
   attributes preserved): 31,304 distinct instruction strings → 28,458 task
   families. Episodes without an instruction (task_index 31051, 43,133
   episodes) are grouped by their labeled verb+object chain instead.
2. **Trajectory signature** — keyframes = gripper open/close events +
   end-effector deceleration points; the keyframe EE positions are
   start-centred and arc-length-resampled to K=8 points → a 24-number
   signature per episode (see `descriptors.npz`).
3. **Duplicate rule** — same family AND mean per-point Euclidean distance
   < **3 cm** AND identical gripper open/close pattern. Leader clustering:
   every member is within 3 cm of its cluster leader, so transitive chains
   cannot glue far-apart trajectories together.
4. **Keep TWO per cluster** (`DEDUP_KEEP=2`) — the cluster **medoid** plus the
   member **farthest from the medoid** (greedy max-min), retaining the
   cluster's residual spread. Clusters of size 2 are kept whole.

## Results (eps = 3 cm, keep = 2)

| metric | value |
|---|---|
| episodes assessed | 94,006 |
| episodes in multi-episode clusters | 8,656 (9.21%) |
| multi-episode clusters | 2,260 (largest: 539 frozen no-instruction episodes) |
| episodes kept (hard dedup) | **89,870** |
| episodes removed | 4,136 (4.40%) |

A stricter keep-one-per-cluster variant is archived as
`report_keep1.json` / `keep_list_keep1.txt` (87,610 kept / 6,396 removed).

## Files

- `episode_weights.jsonl` — one line per episode:
  `{episode_id, task_index, cluster_id, cluster_size, weight, representative}`.
  `representative: true` ⇔ kept by hard dedup (2 per multi-cluster). Soft
  weights: representatives 1.0, other members 1/cluster_size — use them to
  reweigh instead of dropping.
- `seg_weights.jsonl` — `{seg_id, weight}` for all 588,151 annotated segments
  (weight inherited from the episode; joins `annotations/segments.jsonl`).
- `keep_list.txt` — episode_id per line, the 89,870 kept episodes.
- `report.json` — the stats table above.
- `descriptors.npz` — `eids` (94,006), `tasks`, `descs` (94,006×8×3 float32,
  metres, start-relative), `grips` (gripper pattern strings): everything
  needed to re-run clustering at a different eps / keep policy.
- `keep_list_keep1.txt`, `report_keep1.json` — archived keep-1 variant.