File size: 8,057 Bytes
0acf8ea
 
9bf6a16
8ca373b
a429eb8
9bf6a16
8ca373b
 
 
 
 
9bf6a16
8ca373b
 
 
 
 
 
 
 
2b46508
 
 
0acf8ea
8ca373b
25550a3
8ca373b
 
 
a429eb8
 
 
8ca373b
9bf6a16
8ca373b
a429eb8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8ca373b
176740c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8ca373b
 
9bf6a16
8ca373b
9bf6a16
8ca373b
9bf6a16
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
---
license: other
language:
- en
pretty_name: ProcessBench
task_categories:
- visual-question-answering
tags:
- robotics
- embodied-ai
- benchmark
- vision-language-models
- process-understanding
- manipulation
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/processdata_sft.parquet
  - split: eval
    path: data/processbench_eval.parquet
---

# ProcessBench

## Dataset Summary

ProcessBench is a process-aware benchmark for vision-language robotic manipulation understanding. It evaluates whether VLMs can infer how a manipulation execution unfolds, including phase, contact, motion, bimanual coordination, primitive-local progress, temporal order, outcome, and primitive-level transitions.

This release contains 57,892 QA rows: 48,841 SFT rows and 9,051 evaluation rows across 12 task families and 260 manipulation tasks. The split follows strict episode / recording / scene isolation. The benchmark is derived from GM-100, RH20T, REASSEMBLE, and AIST-Bimanual.

## Task Families

ProcessBench contains 12 diagnostic question families:

- **T1 Phase Recognition:** identify the current coarse process phase.
- **T2 Contact Detection:** determine whether task-relevant contact has occurred.
- **T3 Motion Direction Prediction:** infer the dominant motion direction from short temporal context.
- **T4 Bimanual Coordination State:** identify the current coordination state of two arms.
- **T5 Primitive-local Progress:** estimate progress within the current local manipulation step.
- **T6 Motion State Recognition:** distinguish actively moving from stationary states.
- **T7 Operation Outcome Prediction:** predict eventual success or failure from partial execution evidence.
- **T8 Temporal Ordering:** reconstruct the chronological order of shuffled observations.
- **T9 Temporal Priority Prediction:** decide which of two observations occurred earlier.
- **T10 Current Primitive Recognition:** identify the current low-level primitive.
- **T11 Next Primitive Prediction:** infer the next primitive from local process context.
- **T12 Primitive Chain Restoration:** restore a masked primitive in a local primitive chain.


## Repository Contents

This release is organized around a minimal reviewable benchmark package:

- `splits/`: SFT and Eval QA entries.
- `metadata/`: split statistics, task distribution, and license notes.
- `examples/`: rendered representative task cards examples.
- `ProcessData-SFT-Qwen/`: ProcessData-SFT-Qwen LoRA adapter weights and configurations.
- `ProcessData-SFT-Qwen_results/`:  ProcessData-SFT-Qwen predictions and score summaries.
- `benchmark_card.md`: benchmark-level documentation.
- `croissant.json`: Croissant metadata with core and Responsible AI fields.

Full upstream videos and full frame dumps are not redistributed in this release.

## SFT-Eval Split Reconstruction

### Relevant Assets

- Public QA rows for `eval` and `sft`.
- Split manifests and split counts.
- Public release-relative references for each item.

### How to use the public references

Each row exposes three reconstruction-oriented fields:

- `visual_ref`
- `source_episode_ref`
- `reconstruction_key_json`

These are designed to be stable, path-free references. They are sufficient to map a released item back to the corresponding source episode or recording once you have obtained the upstream dataset under its original terms.

### Source-specific units

- `GM-100`: grouped by `(task_id, episode_id)`.
- `RH20T`: grouped by `recording_id`.
- `REASSEMBLE`: grouped by `recording_id`.
- `AIST-Bimanual`: grouped by `recording_id`.

### Expected workflow

1. Obtain upstream datasets under their original terms.
2. Use `source`, `source_task_id`, `source_unit_id`, `camera`, and `frame_indices_json` from the public release.
3. Reconstruct the required visual inputs with the source-specific evaluation scripts in the main codebase.


## Public QA fields Schema

- `item_id`: Stable public item identifier.
- `split`: `eval` or `sft`.
- `source`: `GM-100`, `RH20T`, `REASSEMBLE`, or `AIST-Bimanual`.
- `source_slug`: Short source slug used in release-relative references.
- `source_task_id`: Source-side task identifier without local paths.
- `source_unit_type`: Split-isolation unit type, e.g. `episode` or `recording`.
- `source_unit_id`: Public unit identifier without local paths.
- `task_id`: Canonical paper task ID `T1` ... `T12`.
- `task_name`: Canonical paper task name.
- `task_type_legacy`: Legacy engineering task name when it differs from paper naming.
- `input_type`: Public input format description.
- `question`: Public question text.
- `choice_A` ... `choice_F`: Flattened answer choices.
- `answer`: Public answer label in the release schema.
- `answer_text`: Public answer text mapped from `answer`.
- `num_choices`: Number of non-null answer choices.
- `num_frames`: Number of referenced frames.
- `frame_indices_json`: JSON-encoded frame indices.
- `display_labels_json`: JSON-encoded panel labels such as `["X", "Y"]` or `["X", "Z", "Y"]`.
- `camera`: Source camera identifier when available.
- `arm_type`: Source-side arm configuration tag when available.
- `visual_ref`: Release-relative visual reference string. This is not a local filesystem path.
- `source_episode_ref`: Release-relative source unit reference.
- `reconstruction_key_json`: Minimal JSON payload needed to reconstruct the sample with upstream data and source-specific scripts.
- `task_meta_in_source`: Whether source-side task metadata/context exists upstream.
- `task_meta_public`: Always `false` in this release build. Separate structured task-meta fields are intentionally withheld.
- `split_version`: Current split version.
- `split_group_id`: Source-side split group identifier.
- `builder_version`: Public builder version tag.
- `prompt_version`: Public prompt-serialization version tag.
- `sft_target`: SFT-only supervised target string. Present only in `processdata_sft.*`.

### Notes

- `T5` is equivalent to `T_progress`.
- `T8` is equivalent to `T_temporal`.
- `T9` is equivalent to `T_binary`.
- `T8` is normalized into six explicit permutation choices so the public table remains flat and Parquet-friendly.
- `T9` is normalized to `A/B` choices even when the source data originally used `X/Y`.



## Prompt Templates

This file documents the public prompt serialization used by the current release build. Separate source-side task-meta prompting is intentionally disabled in this public release.

### Standard multiple-choice template

```text
{question}

A: {choice_A}
B: {choice_B}
...

Choose exactly one option label.

Output protocol:
- You may include brief reasoning before the final answer.
- The final line must be exactly: <ANSWER>A</ANSWER>
- Do not output anything after </ANSWER>.
```

### `T8` Temporal Ordering template

```text
You are shown 3 frames from a robot manipulation task.
The frames are labeled X, Y, Z (these labels are arbitrary identifiers, not positional, and not time-ordered).
Determine the correct chronological order of these frames (from earliest to latest).
Choose exactly one 3-letter permutation, for example: YXZ means Y happened first, then X, then Z.

Output protocol:
- You may include brief reasoning before the final answer.
- The final line must be exactly: <ANSWER>XYZ</ANSWER>
- Do not output anything after </ANSWER>.
```

### `T9` Temporal Priority template

```text
A single comparison image shows two labeled robot-manipulation panels from the same episode.
The left-right placement of the panels and the labels are arbitrary identifiers and do not indicate temporal order.
Which labeled panel happened earlier in the real manipulation sequence, X or Y?
Choose exactly one label: X or Y.

Output protocol:
- You may include brief reasoning before the final answer.
- The final line must be exactly: <ANSWER>X</ANSWER>
- Do not output anything after </ANSWER>.
```


## License and Terms

This release uses `license: other`.

- Derived benchmark metadata in this release remains subject to upstream dataset terms.