---
license: other
pretty_name: P&W subtask benchmark
---

# P&W Subtask Segmentation and Failure Detection Benchmark

We built a model that automatically segments P&W rollouts into subtasks and detects failed attempts, for each of P&W's five benchmark tasks.

**Input** (one episode). It needs no prompt and no per-episode human input.
- The three synchronised camera streams: static, left wrist and right wrist.
- The robot's recorded state: gripper and arm signals, logged with each episode.
- The task's subtask list.

**Output** (per episode), in P&W's own label format.
- The full subtask sequence, as contiguous segments with start and end times and labels from the task's vocabulary.
- For each attempt, success or failure, with a probability of success.

**How it scales.**
- One model per task. Training a task needs only that task's labelled episodes (demos and rollouts) and runs on a single GPU.
- Adding a task, or switching to a different labelling convention, means retraining on labelled episodes, not new engineering.
- Once trained, it labels new episodes automatically, with no per-episode prompting or human review.
- It is far cheaper to run than a VLM annotator. Our model is small, runs on a single GPU and makes no API calls, while the VLM baselines below process each episode as a long multimodal prompt (about half a million tokens per episode for Terra, about three quarters of a million for Astra).

This repository lets you check the model against **P&W's own subtask labels** (the raw catalog labels) in two cases and compare it with a baseline: **Terra** (GPT-6.1 Terra, a state-of-the-art vision-language model, VLM) used zero-shot. Every number here can be reproduced with a standalone grader you can run yourself.

**Gold labels.** Every score uses P&W's latest published labels, for the rollouts of the five PAW-GEN-10 policies (DiT-Flow, GR00T N1.7, MolmoAct 2, pi0.5 and Rho).

## Headline

All five tasks (mean of the per-task scores), on the same 899 test rollouts, against P&W's labels. Segment F1 and failure F1: higher is better. Boundary error: lower is better. Tables in this README show point estimates; 95% confidence intervals for every number are in `results/*.json`.

| | Episodes scored | Segment F1 (IoU 0.8) | Mean boundary error | Failure F1 |
|---|---|---|---|---|
| Ours, new rollouts of a known policy family | 899 test rollouts | 87.6 | 0.25 s | 82.1 |
| Ours, a policy family not seen in training | 899 test rollouts, each family held out in turn | 81.4 | 0.31 s | 74.9 |
| Terra zero-shot | 899 test rollouts | 73.2 | 0.40 s | 70.4 |

## Key findings

- **On a known policy family, ours is ahead of Terra zero-shot on 13 of 15 task–metric pairs and level on DC jack failure F1 and pour segmentation.** Averaged over tasks: segment F1 87.6 vs 73.2, boundary error 0.25 s vs 0.40 s, failure F1 82.1 vs 70.4. The gap widens as the overlap threshold gets stricter.
- **The biggest gains are on toolbox and sort, the tasks with many similar or repeated steps** (toolbox +27.3 and sort +21.4 segment F1), and in failure detection (about +9 to +18 on sort, pour, cable and toolbox).
- **A new policy family costs about 6 points of segment F1 and 7 of failure F1 on average.** A held-out family is still ahead of Terra on the mean of all three metrics (per task: ahead on 9 of 15 task–metric pairs, level on 6, behind on none), and adding the family's rollouts to training recovers most of the drop.
- **Segmentation needs little labelled data.** With 0.5 h of labelled video per task, segment F1 is ahead of Terra zero-shot on four tasks and about level on pour. About 4 h gets within 2 points of the full result on average. Failure detection keeps improving with more data, up to the full set.
- **Label consistency lowers every predictor's measured score.** Part of every predictor's measured error comes from inconsistencies in the reference labels (Appendix B). The model can be retrained on whichever convention you settle on.

## What's included

| Folder | Contents |
|---|---|
| [`gold/rollouts/<task>.jsonl`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/tree/main/gold/rollouts) | P&W's latest published subtask labels for the 4,499 PAW-GEN-10 rollouts (5 policies × 900, minus one Rho toolbox rollout excluded because its labels are not contiguous), with `split` (`test`: 899; `cross-validation`: 3,600) and `policy` fields. |
| [`gold/demos/<task>.jsonl`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/tree/main/gold/demos) | P&W's raw subtask labels for the 1,547 demos, included for completeness (the models train on them). |
| [`predictions/seen-policy/<task>.jsonl`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/tree/main/predictions/seen-policy) | Our predictions for every rollout, in P&W's format (`startSec`, `endSec`, `label`, `rating` ∈ failed/excellent, `pSuccess`). Test rollouts come from models that never saw them. Cross-validation rollouts are predicted out of fold (5 folds by object placement). |
| [`predictions/new-family/`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/tree/main/predictions/new-family) | Predictions for every rollout with its policy family held out of training (`held_out_family` field). |
| [`baselines/terra-zero-shot/`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/tree/main/baselines/terra-zero-shot) | Terra's zero-shot labels for the 899 test rollouts (see "Terra baseline" below). |
| [`predictions/train-mix/`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/tree/main/predictions/train-mix) | Predictions from an earlier comparison of training mixes (rollouts only, demos only, demos + rollouts; 3 seeds), made before the latest labels. It used an earlier episode set that predates the current split (its figure captions say 1,074 test rollouts). |
| `results/train-mix.json`, `results/train-mix-grid.json`, `results/train-mix-budgets.json` | Scores of that training-mix comparison, on the same earlier episode set. |
| `plots/train-mix-*.svg` | Figures of that training-mix comparison, on the same earlier episode set. |
| [`grader/grade.py`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/blob/main/grader/grade.py), [`grader/test_grade.py`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/blob/main/grader/test_grade.py) | Standalone grader (Python 3.9+, numpy only) and its unit tests. |
| [`results/`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/tree/main/results) | Every number in this README, with 95% confidence intervals, as the grader's JSON output. |
| [`plots/`](https://huggingface.co/datasets/macrodata/pnw-subtask-benchmark/tree/main/plots) | The figures below. |

Each line is one episode, keyed by `source` + `episode` exactly as in P&W's catalog.

## Two cases

1. **New rollouts or checkpoints of a known policy family** ("seen policy"). The model has been trained on demos and on labelled rollouts of the same policy families, and is applied to new rollouts. This is the everyday case: a new evaluation round of policies you already benchmark. Results: `predictions/seen-policy/`.
2. **A new policy family.** The model has seen no rollouts at all from the policy family it is labelling (leave one family out). This is the case of a new paper or a new architecture entering the benchmark. Results: `predictions/new-family/`.

## Terra baseline

Terra labelled each test rollout zero-shot. It was given:
- the three camera videos and the robot's recorded trajectory (gripper opening and gripper positions);
- the task description and subtask list;
- written labelling instructions.

It saw no labelled examples. We report its first pass, with no review or correction. It covers all 899 test rollouts. 24 of its 4,408 segments have an outcome Terra marked as not visible; the grader counts these as not failed.

We also ran Terra few-shot on 720 of the test rollouts, the four families other than Rho (the few-shot run predates Rho's rollouts, so Rho's 179 test rollouts are left out): the same inputs plus three worked, labelled examples per task. It changed little. Paired on the same episodes, segment F1 rose 2.5 points (74.6 to 77.2, mostly on sort), failure F1 by 2.3 (70.1 to 72.4, not significant) and boundary error stayed at 0.40 s, still well behind our model. We report the zero-shot run as the baseline (`results/terra-few-shot.json`). We also ran Astra (GPT-6 Astra, a larger and slower VLM) few-shot; it gave no meaningful improvement either.

## Metrics and how to run the grader

```bash
uv run grader/grade.py --gold gold/rollouts --pred predictions/seen-policy --split test
uv run grader/grade.py --gold gold/rollouts --pred predictions/seen-policy --pred baselines/terra-zero-shot --split test --common
```

The grader needs only Python and numpy, and prints its results as JSON: point estimates, 95% confidence intervals and, when several predictions are given, paired differences against the first. Its docstring (`uv run grader/grade.py --help`, or the top of the file) documents every option, the matching rules and the output format.

- **Segment F1 @ IoU 0.8.** A predicted segment counts as correct if it has the right label and overlaps the true segment enough. Overlap is measured as intersection over union (IoU): the time the two share, divided by the time they cover together. We use 0.8 in the tables because it is strict enough that boundary placement matters, not just finding the right steps.
- **Mean boundary error (s).** How far, on average, predicted subtask boundaries are from the true ones (boundaries more than 2 s away count as unmatched). Boundary recall and boundary precision give the share of true and of predicted boundaries that are matched.
- **Failure F1.** Per attempt, with each retry its own attempt: how well failed attempts are found. `ok` and `excellent` both count as success.
- **Uncertainty.** Every score comes with a 95% confidence interval: the range the score would likely fall in on a different sample of episodes of the same kind. We estimate it the standard way, by recomputing the score on many random re-draws of the test episodes. The intervals are in `results/*.json` and shown as error bars in the figures. In comparison tables, **bold** marks the better value; the text says where a difference is within noise. The all-tasks row is the unweighted mean of the five tasks.

## Results

### Case 1 — known policy family, test rollouts (899)

Ours vs Terra zero-shot on the 899 test rollouts, scored against P&W's labels. Segment F1 and failure F1: higher is better. Boundary error: lower is better. Each task's result comes from a single model (one training seed), so the confidence intervals reflect episode sampling only, not training-seed variation.

| Task | Seg F1@0.8 (Ours / Terra) | Boundary err, s (Ours / Terra) | Failure F1 (Ours / Terra) |
|---|---|---|---|
| Connect DC jack | **84.7** / 75.0 | **0.35** / 0.61 | **84.2** / 79.8 |
| Sort screws into bins | **80.5** / 59.1 | **0.26** / 0.35 | **85.5** / 76.2 |
| Pour screws into funnel | **95.0** / 94.1 | **0.19** / 0.37 | **83.9** / 66.0 |
| Route cable through closed hoops | **88.3** / 75.3 | **0.25** / 0.32 | **78.5** / 67.6 |
| Open toolbox with screwdriver | **89.5** / 62.2 | **0.22** / 0.37 | **78.5** / 62.7 |
| **All tasks (mean)** | **87.6** / 73.2 | **0.25** / 0.40 | **82.1** / 70.4 |

Ours is ahead on 13 of the 15 task–metric pairs, each with a paired 95% interval that excludes zero; it is level on DC jack failure F1 and pour segmentation. On average the model is ahead by 14.5 points of segment F1 and 11.7 points of failure F1, and its boundary error is 37% lower (0.25 s vs 0.40 s).

![Known policy family, test](plots/metrics-test.svg)

*Figure 1. Segment F1 at IoU 0.8, mean boundary error and failure F1 per task on the 899 test rollouts, ours (known policy family) vs Terra zero-shot. Error bars: 95% confidence intervals.*

Our failure precision and recall, and boundary recall and precision (share of P&W's boundaries, and of ours, matched within 2 s), on the same 899 test rollouts. Higher is better for all four.

| Task | Failure precision | Failure recall | Boundary recall | Boundary precision |
|---|---|---|---|---|
| Connect DC jack | 83.8 | 84.5 | 93.3 | 91.2 |
| Sort screws into bins | 85.3 | 85.7 | 96.1 | 88.1 |
| Pour screws into funnel | 89.9 | 78.7 | 96.8 | 99.1 |
| Route cable through closed hoops | 77.4 | 79.6 | 94.9 | 90.5 |
| Open toolbox with screwdriver | 78.2 | 78.9 | 98.6 | 97.8 |
| **All tasks (mean)** | 82.9 | 81.5 | 95.9 | 93.3 |

Precision and recall are balanced (82.9 vs 81.5 on average). Recall is lowest on pour, cable and toolbox (79–80%): about 1 in 5 of their failed attempts is missed. Boundary recall is 93–99% and boundary precision 88–99% on every task, so nearly every labelled boundary has a predicted one within 2 s, and few predicted boundaries are spurious.

**Probability of success.** Most attempts get a probability close to 0% or 100%, and for those it is reliable. In between it is rougher: low probabilities overstate the chance of success, and high ones slightly understate it. The full reliability table is in `results/calibration.json`.

**Cross-validation.** On the 3,600 cross-validation rollouts (predicted out of fold), the same recipe scores segment F1 86.4, boundary error 0.27 s and failure F1 80.8 (`results/cross-validation.json`).

### Case 2 — new policy family

Each policy family is held out of training in turn. Its rollouts are predicted by a model trained on everything else: demos and the other families' rollouts. The families are the five PAW-GEN-10 policies.

**What a new family costs.** The table pairs each family's test rollouts: the Case 1 models (family seen in training) against the model with that family held out, on the same episodes. Scores are means over the five tasks.

| Policy family | Test rollouts | Seg F1@0.8: seen → held out (change) | Failure F1: seen → held out (change) | Boundary err (s): seen → held out (change) |
|---|---|---|---|---|
| DiT-Flow | 180 | 88.9 → 83.7 (−5.2) | 85.2 → 82.0 (−3.2) | 0.28 → 0.30 (+0.03) |
| GR00T N1.7 | 180 | 84.9 → 80.0 (−4.9) | 81.9 → 73.8 (−8.2) | 0.30 → 0.33 (+0.03) |
| MolmoAct 2 | 180 | 92.5 → 87.5 (−5.0) | 80.8 → 72.3 (−8.5) | 0.24 → 0.28 (+0.04) |
| pi0.5 | 180 | 88.6 → 84.0 (−4.6) | 79.0 → 70.9 (−8.2) | 0.24 → 0.28 (+0.04) |
| Rho | 179 | 85.6 → 75.8 (−9.9) | 84.5 → 80.2 (−4.3) | 0.22 → 0.33 (+0.11) |

Holding out a policy family costs about 6 points of segment F1 and 7 of failure F1 on average (87.6 → 81.4 and 82.1 → 74.9), with boundary error 0.25 → 0.31 s. Most families lose 4.6–5.2 points of segment F1; Rho loses most, 9.9 points on the family mean, and within it DC jack drops 19.4. Each family has only about 180 test rollouts, so these per-family changes have wide intervals (`results/families.json`). Against Terra, a held-out family stays ahead on the mean of all three metrics; per task, it is ahead on 9 of the 15 task–metric pairs, level on 6 and behind on none (below).

**Against Terra.** Held-out-family model vs Terra zero-shot on the same 899 test rollouts:

| Task | Seg F1@0.8 (New family / Terra) | Boundary err, s (New family / Terra) | Failure F1 (New family / Terra) |
|---|---|---|---|
| Connect DC jack | **77.5** / 75.0 | **0.43** / 0.61 | **83.9** / 79.8 |
| Sort screws into bins | **71.7** / 59.1 | **0.30** / 0.35 | **81.5** / 76.2 |
| Pour screws into funnel | 91.1 / **94.1** | **0.23** / 0.37 | **71.3** / 66.0 |
| Route cable through closed hoops | **82.9** / 75.3 | **0.30** / 0.32 | **73.2** / 67.6 |
| Open toolbox with screwdriver | **83.7** / 62.2 | **0.27** / 0.37 | **64.7** / 62.7 |
| **All tasks (mean)** | **81.4** / 73.2 | **0.31** / 0.40 | **74.9** / 70.4 |

Six of these differences are within noise (paired 95% interval includes zero): DC jack segment and failure F1, toolbox and pour failure F1, pour segment F1 (Terra 94.1 vs 91.1, interval includes zero) and cable boundary error (`results/new-family-test.json`).

![New family](plots/metrics-new-family.svg)

*Figure 2. The same metrics when the policy family is held out of training, vs Terra zero-shot, on the 899 test rollouts. Error bars: 95% confidence intervals.*

### Segment F1 and the overlap threshold

A predicted segment counts as correct only if it has the right label and overlaps the matching true segment by at least the threshold, measured as IoU (shared time ÷ combined time). At 0.1, a segment that roughly lands in the right place counts even if its start and end are several seconds off. At 0.8, its boundaries must closely match the true ones: for a 10 s segment, the start and end errors can add up to only about 1–2.5 s, depending on whether the prediction is shifted, too short or too long. So low thresholds mostly test whether the right steps are found in the right order, and high thresholds also test boundary precision. That is why every score falls as the threshold rises, and why the gap to Terra opens up at strict thresholds.

Segment F1, mean of the five tasks, 899 test rollouts:

| IoU threshold | 0.1 | 0.3 | 0.5 | 0.7 | 0.8 | 0.9 |
|---|---|---|---|---|---|---|
| Ours, known family | 96.6 | 96.1 | 94.9 | 92.3 | 87.6 | 75.0 |
| Ours, new family | 95.0 | 94.3 | 92.1 | 87.2 | 81.4 | 66.7 |
| Terra zero-shot | 93.3 | 92.3 | 89.0 | 81.4 | 73.2 | 50.7 |

In both cases, the gap to Terra is smallest at lenient thresholds and largest at the strictest. At IoU 0.1 the known-family model leads Terra by 3.3 points on the mean. At 0.8 the lead is 14.5 points, and at 0.9 it is 24.3. On pour the two are within 2.2 points up to IoU 0.8; at 0.9 the model leads by 35 points. Per-task values: `results/iou-sweep.json`.

![IoU sweep](plots/iou-sweep.svg)

*Figure 3. Segment F1 against the IoU threshold on the 899 test rollouts, per task and mean: ours with the policy family seen (known family) and held out (new family), and Terra zero-shot. The dashed line marks 0.8, the threshold used in the tables.*

### Segment F1 and a boundary tolerance in seconds

Here a predicted segment counts as correct if its label is right and its start and end are each within T seconds of the true segment's start and end. This is easier to read than IoU because it does not depend on segment length: a 3 s segment and a 30 s segment get the same allowance. Run it with `--seg-tol T`.

Segment F1, mean of the five tasks, 899 test rollouts:

| Tolerance (s) | 0.5 | 1 | 1.5 | 2 | 2.5 | 3 |
|---|---|---|---|---|---|---|
| Ours, known family | 76.6 | 87.8 | 90.6 | 91.7 | 92.7 | 93.4 |
| Ours, new family | 66.7 | 81.4 | 86.0 | 87.3 | 88.8 | 90.2 |
| Terra zero-shot | 52.8 | 70.6 | 77.3 | 80.4 | 84.0 | 85.9 |

With 3 s of slack the known-family model leads Terra by 7.5 points on the mean. The lead grows as the tolerance tightens, to 17.2 points at 1 s and 23.8 at 0.5 s, and with the policy family held out the model still leads by 10.8 and 13.9 points there. On pour the two stay within 2.2 points down to 1 s. Per-task values for all three predictors, with paired 95% CIs: `results/seconds-sweep.json`.

![Seconds-tolerance sweep](plots/seconds-sweep.svg)

*Figure 4. Segment F1 against the boundary tolerance T on the 899 test rollouts, per task and mean: ours with the policy family seen (known family) and held out (new family), and Terra zero-shot. A segment counts if its label is right and its start and end are each within T seconds of the true ones.*

## Per-task notes

- **Connect DC jack.** Boundary error is about half Terra's (0.35 s vs 0.61 s) and segment F1 is 84.7 vs 75.0; failure F1 is level (84.2 vs 79.8).
- **Sort screws into bins.** The hardest task for segmentation, for both predictors (80.5 vs 59.1): it has many short pick/place cycles. Screw colour is not scored: the black- and silver-screw pickups count as one label, and our predictions do not name the colour. At the released camera resolution (320×240 for the wrist cameras) and video compression the two colours are often hard to tell apart, and P&W's own labels get it wrong in places: in at least 137 pickups across 89 rollouts, the labelled colour disagrees with the screw the gripper closed on, as inferred from the recorded gripper position and the initial screw layout. Scoring colour would mostly measure this ambiguity rather than the segmentation. `--keep-colour` scores the colours separately.
- **Pour screws into funnel.** Segmentation is level (95.0 vs 94.1; three subtasks), but failure detection separates the two: 83.9 vs 66.0, mostly through precision (89.9 vs 58.1).
- **Route cable through closed hoops.** Segment F1 88.3 vs 75.3. A re-grasp after a failed insertion is labelled both ways in P&W's labels (Appendix B), which counts against both predictors.
- **Open toolbox with screwdriver.** The largest segmentation gap (89.5 vs 62.2); failure F1 78.5 vs 62.7.

## Data efficiency

How much labelled video does a new task need? We retrain on nested subsets of the training set. Hours count all labelled training video, demos plus rollouts. Test rollouts are never part of training.

Each cell gives segment F1@0.8 / failure F1 on the 899 test rollouts against P&W's labels (higher is better for both). Values are means over 3 seeds. The "all data" column gives in brackets the size of each task's full training set, which differs per task; pour's is under 4 h, so its 4 h subset is the full set. Subsets are nested and made of whole object placements. "Maximum data" adds P&W's Research Replication rollouts (on training placements) to all data.

| Task | 0.5 h total | 1 h total | 2 h total | 4 h total | all data | maximum data |
|---|---|---|---|---|---|---|
| Connect DC jack | 80.4 / 70.2 | 81.8 / 78.0 | 83.3 / 82.2 | 83.8 / 82.8 | 84.6 / 85.0 (12.1 h) | 85.2 / 85.4 (16.4 h) |
| Sort screws into bins | 65.5 / 71.4 | 68.8 / 75.3 | 72.2 / 79.2 | 73.8 / 81.3 | 80.3 / 85.5 (14.4 h) | 80.4 / 85.6 (15.2 h) |
| Pour screws into funnel | 93.5 / 68.1 | 93.6 / 76.3 | 94.7 / 78.7 | = all data | 95.4 / 83.2 (3.3 h) | 94.3 / 81.5 (6.4 h) |
| Route cable through closed hoops | 83.1 / 66.2 | 84.0 / 69.3 | 85.0 / 72.6 | 87.0 / 76.0 | 88.3 / 78.7 (11.9 h) | 87.9 / 78.0 (12.3 h) |
| Open toolbox with screwdriver | 80.2 / 61.9 | 85.4 / 65.1 | 87.3 / 67.3 | 88.0 / 72.4 | 89.2 / 78.9 (9.2 h) | 89.8 / 79.6 (10.8 h) |

With 0.5 h, segment F1 is already ahead of Terra's zero-shot test score on DC jack, sort, cable and toolbox (80.4 vs 75.0, 65.5 vs 59.1, 83.1 vs 75.3, 80.2 vs 62.2) and about level on pour (93.5 vs 94.1). These are 3-seed means, not a paired test.

At 4 h, segment F1 is on average 2.0 points below the full-data result; sort is furthest, at 6.6. Failure F1 gains more from data: on average +14.7 points from 0.5 h to all data, against +7.0 for segment F1. It is still rising between 4 h and the full set (DC jack +2.2, sort +4.2, cable +2.7, toolbox +6.5). Adding the Research Replication rollouts changes no score by more than 1.7 points, and slightly lowers pour (94.3 / 81.5 at maximum data vs 95.4 / 83.2 with all data).

![Data efficiency](plots/data-efficiency.svg)

*Figure 5. Segment F1 at IoU 0.8, mean boundary error and failure F1 on the 899 test rollouts, against hours of labelled training video per task (demos + rollouts, log scale). Mean ± sd over 3 seeds; the rightmost line point uses all training data, the open diamond the maximum-data point.*

## Conclusions

- **On a known policy family, ours is ahead of Terra zero-shot on 13 of 15 task–metric pairs and level on DC jack failure F1 and pour segmentation.** Averaged over tasks: segment F1 87.6 vs 73.2, boundary error 0.25 s vs 0.40 s, failure F1 82.1 vs 70.4. The gap widens as the overlap threshold gets stricter.
- **The biggest gains are on toolbox and sort, the tasks with many similar or repeated steps** (toolbox +27.3 and sort +21.4 segment F1), and in failure detection (about +9 to +18 on sort, pour, cable and toolbox).
- **A new policy family costs about 6 points of segment F1 and 7 of failure F1 on average.** A held-out family is still ahead of Terra on the mean of all three metrics (per task: ahead on 9 of 15 task–metric pairs, level on 6, behind on none), and adding the family's rollouts to training recovers most of the drop.
- **Segmentation needs little labelled data.** With 0.5 h of labelled video per task, segment F1 is ahead of Terra zero-shot on four tasks and about level on pour. About 4 h gets within 2 points of the full result on average. Failure detection keeps improving with more data, up to the full set.
- **Label consistency lowers every predictor's measured score.** Part of every predictor's measured error comes from inconsistencies in the reference labels (Appendix B). The model can be retrained on whichever convention you settle on.

## Appendix A — labelling conventions that matter for scoring

A few convention choices move every predictor's scores, because they decide how many segments an episode has and where their boundaries sit. P&W's current labels do not always make the same choice. The choices are:

1. **Failed or empty attempts.** A failed grasp followed by moving on: label it as the next step (Pour, Place) or keep it as the pick-up?
2. **Retries.** Each new attempt (a new grasp, insertion or pry) can be its own segment with its own rating, or merged into the previous attempt.
3. **Going back.** When the robot returns to an earlier subtask, for example re-grasping a dropped cable, that can be labelled as a new segment for the earlier subtask, or left inside the later one.
4. **Boundary anchors.** Boundaries can be tied to a physical event (for example the start of the approach for the next subtask, or the moment an object leaves the table) or placed by eye.
5. **Object identity.** Which connector, screw colour, hoop or latch a segment refers to can be taken from what the wrist cameras show at the time, or from the usual task order.

Once a convention is fixed, the model can be retrained on it.

## Appendix B — label consistency notes

Below are examples where P&W's labels for similar situations differ. The times are P&W's labels, in seconds. Some examples come from P&W's published Research Replication rollouts, which are not part of the benchmark set; the counts are over the benchmark's PAW-GEN-10 rollouts. Statements about what happens on screen are marked "in the video"; we checked each one against the footage.

- [`d10-interpolation/molmoact2-policy/route-cable-through-closed-hoops/27`](https://pokeandwiggle.com/leaderboard/molmoact-2/d10-interpolation/route-cable-through-closed-hoops?episode=27). Labels: Pick up the cable 0–5.71 s (excellent), then two failed first-hoop insertions, 5.71–15.28 s and 15.28–27.83 s, with no Pick up between them. In the video, the cable drops from the gripper at about 14 s and is grasped again by about 17 s, before the second insertion. Across the 900 cable rollouts, a failed insertion is followed by a new Pick up in 28 cases and directly by another insertion in 26.
- [`cupid-iteration-2/method/pour-screws-into-funnel/28`](https://pokeandwiggle.com/research-replication/cupid/iterations/2/pour-screws-into-funnel/method?episode=28). Labels: Pour 0–2.88 s, rated excellent, then Place the container back 2.88–12.97 s, rated failed. There is no Pick up segment. None of the 900 pour rollouts in the benchmark has labels that start with Pour. In the video, the gripper misses the container's edge and the container stays on the table.
- [`d10-interpolation/molmoact2-policy/pour-screws-into-funnel/15`](https://pokeandwiggle.com/leaderboard/molmoact-2/d10-interpolation/pour-screws-into-funnel?episode=15). Labels: Pick up the container 0–2.83 s (failed), then Pour 2.83–6.93 s (failed). In the video, the gripper rises empty and reopens near the container rather than moving towards the funnel. Across the pour rollouts, a failed pick-up is followed by Pour in 38 cases, by another pick-up in 136, and ends the episode in 245.
- [`dsd-iteration-1/baseline/open-toolbox-with-screwdriver/23`](https://pokeandwiggle.com/research-replication/dsd/iterations/1/open-toolbox-with-screwdriver/baseline?episode=23). Labels: one Push segment for the whole episode, 0–35.87 s, rated failed (28 of the 899 toolbox rollouts are a single Push segment). In the video, the left gripper works at the first latch with its fingers at about 24–27 s, while the screwdriver stays on the table; no Release segment is labelled.
- [`d10-nominal/molmoact2-policy/connect-dc-jack/09`](https://pokeandwiggle.com/leaderboard/molmoact-2/d10-nominal/connect-dc-jack?episode=9). Labels: one segment, Pick up the first female DC jack connector, 0–8.53 s, failed. In 898 of the 900 DC jack rollouts the labels start with the male connector. In the video, the right wrist camera shows the male tip at about 3–8.5 s.
- [`open-loop-execution-iteration-1/baseline/route-cable-through-closed-hoops/07`](https://pokeandwiggle.com/research-replication/open-loop-execution/iterations/1/route-cable-through-closed-hoops/baseline?episode=7). Labels: Pick up the cable 0–7.03 s (excellent), then one failed first-hoop insertion, 7.03–37.23 s. No segment names the second hoop. In the video, the robot passes the first hoop and inserts the cable end into the second hoop at about 30–37 s.

All the scores in this benchmark are measured against P&W's labels as they are, so differences like these count as errors for every predictor.