---
title: "Segment-aware ISR: giving ISR task semantics with a VLM — does it beat plain ISR?"
subtitle: "Experiment 3 · ALOHA transfer-cube · 50 human demos · ACT · 3 seeds × 200 rollouts · 2026-09-16"
---

# 1. Why

Plain ISR resamples every episode with one rule — keep a frame every `d` degrees of joint motion — regardless of what the robot is doing. It cannot tell a deliberate steadying pause before a grasp from idle hesitation, and it samples a free-space transit the same way as a handover. Experiment 2 showed plain ISR at ⅓ of the frames reaches 30 % success where full data reaches 60 %.

Hypothesis: if a VLM tells us which skill phase each frame belongs to and which pauses are deliberate, per-phase knobs (dense where precision matters, sparse in transit, pauses preserved) should train a better policy than plain ISR.

Related: ESPADA (arXiv 2512.07371) segments demos with a VLM and downsamples non-critical parts, but compares to heuristic downsampling, not to ISR.

# 2. Pipeline

**Segmentation (verified).** Two Gemini 2.5 Pro passes per episode via OpenRouter:
1. *Video pass* — 25 fps clip + kinematic summary (speed bins, gripper events, quiet holds) → skill labels and a deliberate/incidental verdict for each kinematically detected hold (69 holds; 44 deliberate).
2. *Grid pass* — 33 timestamp-burned frames → event times: grasp close, lift, handover contact, right release.

Verification: the video pass's boundaries were ~1 s late (checked frame by frame), so **segments are rebuilt from the grid anchors**: transit → approach (last 1 s before grasp) → grasp → carry → handover → retreat. 50/50 episodes contiguous and monotone; one episode (37) needed a re-query. Hold verdicts were checked only statistically (they cluster where expected), not frame by frame.

![](figures/fig_segments_final.png)

**Selection.** The official ISR dynamic program is run *per phase* with a phase-specific `d`; VLM-flagged deliberate holds keep every 3rd frame ("dwell"); gripper events are force-kept. The total kept-% is a *result* of the knobs, chosen from a 150-config lookup table (`exp3/tables/sa_sweep.md`), not a target.

![](figures/fig_keeps_isr_vs_saF.png)

# 3. Arms

| arm | precision phases (approach/grasp/handover) | transit/carry/retreat | deliberate pauses | kept |
|---|---|---|---|---|
| raw | all | all | all | 100 % |
| plain ISR | d 3.5° | d 3.5° | dropped | 32 % |
| sa_A | 1° | 8° | dropped | 33 % |
| sa_D | 1° | 8° | every 3rd | 34 % |
| sa_F | 2° | 4° | every 3rd | 31 % |
| sa_G | 1.5° | 6° | every 3rd | 31 % |
| sa_H | 1° | 3.5° (= ISR) | every 3rd | 41 % |
| sa_E | 0.5° | 8° | every 3rd | 50 % |
| plain ISR @ 50 % | 1.75° | 1.75° | dropped | 52 % |

Same policy and eval as Experiment 2: LeRobot ACT (chunk 100, batch 8, 20k steps), seeds 1000/2000/3000, `gym_aloha/AlohaTransferCube-v0`, 200 episodes per checkpoint, playback held to the training-data rate (3 for ⅓ arms, 2 for ~½ arms).

# 4. Results (rate-matched success %, mean ± sd over 3 seeds)

![](figures/exp3_sr_vs_kept.png)

| arm | kept | success % (mean ± sd) | per seed | native-rate |
|---|---|---|---|---|
| raw | 100 % | **60.0 ± 3.6** | 56 / 58 / 65 | — |
| plain ISR (3.5°) | 32 % | **30.2 ± 7.7** | 22 / 40 / 28 | 10.8 |
| sa_A 1°/8°, pauses dropped | 33 % | **15.2 ± 5.2** | 12 / 22 / 10 | 7.3 |
| sa_D 1°/8°, pauses kept | 34 % | **29.2 ± 7.0** | 30 / 37 / 20 | 14.8 |
| sa_F 2°/4°, pauses kept | 31 % | **36.7 ± 5.2** | 34 / 44 / 32 | 13.8 |
| sa_G 1.5°/6°, pauses kept | 31 % | **36.3 ± 8.6** | 24 / 44 / 40 | 12.5 |
| sa_H 1°/3.5°, pauses kept, no cap | 41 % | **38.8 ± 4.4** | 45 / 36 / 36 | 22.5 |
| sa_E 0.5°/8°, pauses kept | 50 % | **47.5 ± 4.1** | 53 / 43 / 46 | 33.2 |
| plain ISR @ 50 % (1.75°) | 52 % | **49.7 ± 9.0** | 59 / 52 / 38 | 28.8 |


Failure phase (600 episodes per arm):

| arm | never reached cube | touched, no lift | lifted, no handover | success |
|---|---|---|---|---|
| raw | 3 % | 14 % | 24 % | **60 %** |
| plain ISR (3.5°) | 11 % | 30 % | 30 % | **30 %** |
| sa_A 1°/8°, pauses dropped | 20 % | 31 % | 33 % | **15 %** |
| sa_D 1°/8°, pauses kept | 15 % | 26 % | 30 % | **29 %** |
| sa_F 2°/4°, pauses kept | 14 % | 26 % | 23 % | **37 %** |
| sa_G 1.5°/6°, pauses kept | 14 % | 23 % | 26 % | **36 %** |
| sa_H 1°/3.5°, pauses kept, no cap | 14 % | 20 % | 26 % | **39 %** |
| sa_E 0.5°/8°, pauses kept | 10 % | 24 % | 19 % | **48 %** |
| plain ISR @ 50 % (1.75°) | 6 % | 24 % | 20 % | **50 %** |


# 5. Findings

1. **Deliberate pauses carry information: +14 pt** (sa_D 29 vs sa_A 15, every seed). Keeping the VLM-flagged holds cuts grasp failures ("touched, no lift").
2. **Starving transit costs as much.** At d = 8° in transit (11 % of those frames), "never reached the cube" doubles (11 → 20 %). Round 1 therefore tied plain ISR.
3. **With transit kept near ISR's density, segment-aware ISR beats plain ISR at equal budget: sa_F 36.7 vs 30.2, higher on every seed;** sa_G 36.3. Gain is in the grasp phase.
4. **Letting the budget float (sa_H, 41 %) gives 38.8 — +8.6 over ISR at +9 pt of frames.** Native-rate success also doubles (22.5 vs 10.8).
5. **sa_E (50 %) reaches 47.5, i.e. 80 % of raw's success with half the frames.** But plain ISR at the same 52 % budget reaches 49.7 ± 9.0 (59 / 52 / 38) — the gain at 50 % is budget, not segmentation; sa_E is merely more consistent across seeds (sd 4 vs 9).
6. **The handover phase is untouched by any arm** (raw's remaining 30-pt margin lives there); denser sampling there did not help.

# 6. Verdict

Segment awareness makes ISR better, modestly and consistently: +6–9 pt success at equal or slightly larger frame budget, driven by preserving deliberate pauses and not over-compressing transit. It does not close the gap to full data. The method is only as good as its segmentation; the budget-constrained framing of the ISR paper hides the trade-off, and the dynamic-budget arm (sa_H) is the practical setting.

# 7. Caveats

- One task, one policy family, 50 demos, sim; ±5 pt seed noise on 3 seeds — the sa_F/sa_H gains are 1.5–2× their standard error.
- Hold verdicts not verified frame by frame; segmentation cost ~$6 in VLM calls per 50 episodes.
- Knobs for F/G/H were chosen after seeing round 1 (A/D); an independent replication on a second task (insertion, 24 % pauses) is the natural next step.

# 8. Artifacts

`Kavin60606/isr-aloha-transfer-cube-experiment/exp3/`: segments (video pass, grid anchors, final), sweep table, keep files, per-cell results, code, figures, this report. Checkpoints under `isr-aloha-act-ckpts/exp3/`.