---
title: "ISR vs uniform downsampling vs full data — closed-loop success on human teleop demos (ALOHA transfer-cube)"
subtitle: "Experiment 2 · 3 arms × 3 seeds × 200 rollouts · 2026-09-14"
---

# 1. Question

The ISR paper (Yang et al., IROS 2026, arXiv:2606.22907) claims that resampling demonstrations uniformly in *information* (motion + acceleration) instead of *time* raises policy success by ~24 pt over 3× time-uniform downsampling, at the same ⅓ frame budget, on real robots. It never compared against the full, untouched data, never ran in simulation, and never used joint space.

We test, on public human teleop data with a fully reproducible closed-loop evaluation:

1. **ISR vs uniform 3×** at equal frame budget — does *which* frames matter?
2. **ISR vs raw** — does compression cost anything against using every frame?
3. **Playback rate** — what happens when a policy trained on compressed frames is executed at the native control rate?

# 2. Data

`lerobot/aloha_sim_transfer_cube_human`: 50 episodes, a human teleoperating two ALOHA arms in MuJoCo, 400 frames per episode at 50 Hz, 14 joint positions, one 480×640 top camera. Task: right arm picks the cube and hands it to the left arm.

The demos carry real operator timing variance:

| metric | value |
|---|---|
| speed variability (std / mean) | 1.01 |
| stop-go segments per 8 s episode | 4.9 |
| idle frames per episode | 20 % (range 7–34 %) |
| handover time across episodes | 3.6 s ± 0.7 s (2.3–6.2 s) |
| high-jerk frames | 19.5 % |

# 3. ISR as applied

Joint space, 12 arm joints in **degrees** (the paper's numbers are end-effector meters and do not transfer; applied literally they keep 93 % of frames and, with the acceleration term at 88 % of the information, degenerate to uniform sampling).

Knobs were chosen from a full lookup table (`tables/transfer_cube/grid.md`: d_target × λ_acc → kept % and pause-frames kept %), following the paper's protocol:

| knob | value | rationale |
|---|---|---|
| frame budget | ⅓ (32 % kept) | same as the paper's baseline |
| d_target | 3.5° | solves for ⅓ budget |
| λ_acc | 0.0004 | acceleration ≈ 20 % of total information |
| gripper force-keep | 2.84° (5 % of range) | paper's rule |

Effect on the data: 20,000 → 6,412 frames; **pause frames kept 4 %** (uniform 3× keeps 34 %); grasp/handover frames force-kept.

![](figures/fig_isr_selection.png)

# 4. Arms and protocol

| arm | rule | frames | kept |
|---|---|---|---|
| raw | all frames | 20,000 | 100 % |
| uni3 | every 3rd frame (paper baseline) | 6,700 | 34 % |
| isr | ISR, d 3.5°, λ 0.0004 | 6,412 | 32 % |

- **Policy:** LeRobot ACT, ALOHA defaults — ResNet-18, transformer dim 512, chunk 100, batch 8, lr 1e-5, VAE on. **20,000 steps**, seeds 1000/2000/3000. Identical for every arm.
- **Eval:** `gym_aloha/AlohaTransferCube-v0`, **200 rollouts** per checkpoint, 400-step budget, success = cube in the left gripper.
- **Playback:** *native* — one policy action per sim step (a ⅓-budget policy moves 3× faster than the demos); *rate-matched* — each action held 3 sim steps, same 400-step budget.

# 5. Results

![](figures/exp2_results_bars.png)

| arm | kept | native SR % | rate-matched SR % | per-seed (native) | per-seed (rate-matched) |
|---|---|---|---|---|---|
| raw | 100 % | **60.0 ± 3.6** | – | 56 / 58 / 65 | – |
| uni3 | 34 % | 2.3 ± 0.2 | **9.8 ± 3.4** | 2 / 2 / 2 | 6 / 14 / 8 |
| isr | 32 % | 10.8 ± 5.7 | **30.2 ± 7.7** | 5 / 18 / 9 | 22 / 40 / 28 |

mean ± sd over 3 seeds; each cell 200 rollouts (±7 pt). 

![](figures/exp2_sr_vs_kept.png)

# 6. Findings

1. **At equal ⅓ budget, ISR beats uniform 3× by 3× — 30 vs 10 % rate-matched, 11 vs 2 % native.** Every seed agrees (22/40/28 vs 6/14/8). Selecting frames by information rather than by the clock is the better way to compress. The paper's core claim reproduces on public human data in joint space.
2. **Neither compressed arm approaches full data.** Raw reaches 60 %. With only 50 demos, throwing away two-thirds of the frames costs 30 pt even with the best selection. Compression is not free; ISR only halves the damage.
3. **Playback rate is decisive.** Compressed-data policies executed at native rate collapse (2 %, 11 %); holding each action 3 steps recovers 10 % and 30 %. ISR outputs carry no timing, so the deployment side must restore it — the paper's real-robot loop did this implicitly.
4. **Reconciling with the earlier run** (same task, 100 rollouts, 60k steps): uniform 3× scored 31/20 %, ISR-⅓ 42/40 %, raw 62 %. With 20k steps and 200 rollouts the compressed arms are lower in absolute terms, the ranking and the ISR:uniform ratio are unchanged. Compressed arms appear to need more training steps than raw to converge — a further cost of compression.

# 7. Verdict

- ISR is a **better compressor** than time-uniform downsampling: same budget, 3× the success.
- ISR is **not a substitute for data**: at ⅓ budget it loses 30 pt to raw on 50 demos. The earlier run showed ISR at 50 % kept can match raw; the knee lies between ⅓ and ½.
- **Deployment must be rate-aware** or any compressed policy fails.
- Acceleration weighting adds little on human data (earlier λ=0 ablation); pause removal is what works.

# 8. Caveats

- One task, one policy family (ACT), 50 demos, sim.
- Fixed action-repeat approximates rate-matching; true per-frame timing is not recoverable from ISR output.
- 3 seeds: the 20-pt ISR–uniform gap is ~2.5× its standard error — solid direction, ±8 pt magnitude.

# 9. Artifacts

Code, tables, per-cell `eval_info.json`, checkpoints and datasets: `Kavin60606/isr-aloha-transfer-cube-experiment` (folder `exp2/`) and `Kavin60606/isr-aloha-act-ckpts` (`exp2/`).