# std_mm_teleop — teleop standardization run on `mm_teleop`

Pipeline: `/workspace/teleop_std_poc` (ISR resample → ActionVariance score → bucket → plots).
Source: `/workspace/mm_teleop` — LeRobot **v3.0**, `so_follower` (SO-101), 100 episodes,
28 948 frames @ 30 fps, 1 task, 6-D state/action in **degrees** (5 arm joints + gripper).

Run date: 2026-08-18. Env: `/workspace/.venv-tsp` (py3.10, numpy 2.2.6, pyarrow 25.0.1, mpl 3.10.9).

## Command

```bash
cd /workspace/teleop_std_poc
TSP_ROOT=/workspace/std_mm_teleop /workspace/.venv-tsp/bin/python run_all.py \
    --local /workspace/mm_teleop --all-episodes \
    --d-target 12 --lambda-acc 0.002 --gripper-threshold 2.0 --baseline 3 \
    --eps 0.5 --max-episode-figs 12
```

## Pipeline changes this run needed (all backward compatible)

| Change | Why |
|---|---|
| `TSP_ROOT` env var in all 5 stages | outputs land here instead of inside the repo |
| `fetch_trajs.py --local <dir>` / `--all` | the POC only read the HF Hub; this dataset is on disk |
| SO-101 6-D layout in `split_state` | 6-D state was falling through to `unknown[6]` → gripper force-keep disabled |
| v3 parquet cache in `_episode_table_v3` | it re-read the 28 948-row concatenated file once per episode (100×) |
| `isr_resample.py --gripper-threshold` | 0.05 is a *normalized* gripper default; this gripper is in degrees |
| `viz.py --max-episode-figs`, wide-figure scaling | 100 bars/figures were unreadable |

## Knob calibration (defaults assume metres — this data is degrees)

Measured over all 100 episodes: per-frame joint step 2.03° mean / 1.50° median;
accel magnitude 267 °/s² mean; gripper span 41.6° per episode.

- `--lambda-acc 0.002` → the acceleration term carries **21 %** of per-frame info
  (paper's 0.01-in-metres is likewise a minority-but-material share). 0.01 here would be 57 % — acc-dominated.
- `--gripper-threshold 2.0` ≈ 5 % of the 41.6° gripper span — the same intent as the 0.05 normalized default.
  Flags 12.5 % of frame pairs as gripper events (0.5 → 21 %, i.e. jitter).
- `--d-target 12` → 32.8 % of frames kept, deliberately matched to the 3× uniform baseline (33.7 %)
  so ISR-vs-uniform is like-for-like. Sweep is in `calibration_sweep.txt`.

## Results

**Compression** — 28 948 → 9 500 frames (32.8 %). Kept ratio varies **25.1 %–44.1 %** per episode
(std 0.040); the 3× baseline is flat at 33.4 % (std 0.001). That content-adaptivity is the ISR claim,
and it reproduces here. 42 frames/episode on average are gripper-forced.

**Scoring** — dataset ActionVariance **raw 89.8 → ISR 94.3 (+5.0 %)**; it *rose* in 76/100 episodes
(mean +5.4 %). This is the opposite of the BridgeData POC (−22 %) and is expected rather than a bug:
removing pause frames removes the *lowest*-variance samples (a held pose has near-identical actions),
so the mean over what remains goes up. The raw→ISR delta is therefore **not** a "pacing noise removed"
readout on this dataset — read the ISR column alone, as a relative ranking.

**Buckets** (percentile mode, P50/P90 on ISR scores): **50 train / 40 review / 10 quarantine**.
Quarantined: `ep7, ep11, ep15, ep42, ep61, ep71, ep83, ep86, ep87, ep92`.
Score spread 71.0 → 115.1 (p50 95.0, p90 106.1); every episode was scoreable (100 % clustered coverage).

## Caveats

1. **Bucket assignment is ε-sensitive on this dataset** (`out/eps_sensitivity.json`). All 100 episodes
   are the same task, so states overlap heavily. Per-episode rank correlation against ε=0.5:
   ε=0.3 → 0.51, ε=0.2 → 0.41, ε=0.1 → 0.19; mean cluster size 90 → 22 → 8.5 → 3.5 points.
   The train/quarantine split would reshuffle materially at a tighter radius — treat the current
   10 quarantines as a shortlist to eyeball, not a verdict, until ε is pinned by a labeled reference.
2. Scores are in **degrees²** and only comparable within this dataset (the scorer is test-grade —
   implemented from paper Eq. 9, no official code exists).
3. Joint-space ISR is the POC's extension beyond the paper (paper operates on EE positions).
4. Only 5 arm joints drive the resampling; the gripper channel drives force-keeping only.

## Outputs

```
trajs/               100 × ep{N}.npz (positions/gripper/actions/timestamps) + index.json
out/isr/             100 × ep{N}.npz (ISR-kept) + ep{N}_baseline3x.npz + stats.json
out/scores_raw.json  ActionVariance before ISR
out/scores_isr.json  ActionVariance after ISR
out/report.json      final artifact — buckets, thresholds, per-episode ISR stats
out/eps_sensitivity.json   ε-radius stability check
out/plots/           compression.png, spacing.png, variance.png, ep{0..11}_isr.png
run.log              full stdout
```

Nothing was written back into `mm_teleop`; it is untouched.

---

## v3.0 dataset build (2026-08-18, after the POC run)

`build_std_dataset.py` materializes the ISR-kept frames as a real LeRobot **v3.0** dataset at
`lerobot_v30/` (223 MB): parquet rows AND the matching video frames, decoded from `mm_teleop`
and re-encoded h264. 100 eps / 9,500 frames / 1 task, loads via `LeRobotDataset`.

Frames are renumbered onto a uniform grid at **10 fps** — the measured effective rate
(30 fps ÷ 3.05x compression) — because ISR spacing is non-uniform in real time and LeRobot
enforces `timestamp ≈ frame_index / fps`. Episode durations stay within ~2% of the real ones.
`--fps 30` rebuilds at the source rate instead.

Env: `/workspace/.venv-lerobot` (lerobot 0.4.4). Comparison against ego: `COMPARISON_ego_vs_std_mm_teleop.md`.