LALABAB commited on
Commit
cb0f4e7
·
verified ·
1 Parent(s): 5ae1ac0

Dataset card

Browse files
Files changed (1) hide show
  1. README.md +227 -0
README.md ADDED
@@ -0,0 +1,227 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ pretty_name: VLA-Bind and Train-Bind
4
+ language:
5
+ - en
6
+ task_categories:
7
+ - robotics
8
+ tags:
9
+ - vision-language-action
10
+ - attribute-binding
11
+ - benchmark
12
+ - libero
13
+ - robosuite
14
+ - manipulation
15
+ size_categories:
16
+ - 100K<n<1M
17
+ configs:
18
+ - config_name: vla_bind_v1
19
+ data_files: benchmark/scenes/vla_bind_v1.parquet
20
+ default: true
21
+ - config_name: sweeps
22
+ data_files: benchmark/scenes/sweeps.parquet
23
+ - config_name: conflict_and_absent_word
24
+ data_files: benchmark/scenes/conflict_and_absent_word.parquet
25
+ - config_name: language
26
+ data_files: benchmark/scenes/language.parquet
27
+ - config_name: train_bind
28
+ data_files: train_bind/data/*.parquet
29
+ - config_name: train_bind_episodes
30
+ data_files: train_bind/episodes.parquet
31
+ ---
32
+
33
+ # VLA-Bind and Train-Bind
34
+
35
+ > Upload in progress. The scene tables and Train-Bind files appear here over the next hours.
36
+
37
+ VLA-Bind is a benchmark for attribute binding in vision-language-action (VLA) policies. It asks one question: when
38
+ an instruction names an object by an attribute ("the red bowl", "the smaller cup"), does the policy pick the object
39
+ the words name, even when other objects share part of the description? Train-Bind is the matching training set of
40
+ scripted demonstrations.
41
+
42
+ This repository holds the data of the paper *Lost Among Look-Alikes: A Benchmark for Attribute Binding in
43
+ Vision-Language-Action Policies* (under review). The trained adapters are in
44
+ [LALABAB/VLA-Bind-adapters](https://huggingface.co/LALABAB/VLA-Bind-adapters). The scene generator and the evaluation code are not
45
+ part of this release yet. This card describes the data well enough to load and score it without them.
46
+
47
+ All results in the paper use **version 1.0**. Fixes will get a new version number, and version 1.0 stays available.
48
+
49
+ ## What is in the release
50
+
51
+ | Folder | What it holds |
52
+ |---|---|
53
+ | `benchmark/scenes/` | One parquet table per group, one row per scene, with its start state |
54
+ | `benchmark/manifests/` | The same scenes in their original form: one JSON file per split and one `.npz` of start states |
55
+ | `train_bind/data/` | Train-Bind demonstrations, one row per frame, with both camera images |
56
+ | `train_bind/episodes.parquet` | One row per demonstration: instruction, stratum, start state, generation attributes |
57
+ | `train_bind/meta/` | Generation settings per file and the balance report |
58
+ | `train_bind/instruction_rewrites/` | The rewritten instruction of every demonstration used by the LANG rows, and the word lists behind them |
59
+
60
+ ## The benchmark
61
+
62
+ ### Scene families
63
+
64
+ A family is a set of scenes built from one table state, four in the basic case:
65
+
66
+ - **base**: the words name one object;
67
+ - **word swap**: the words change so that they name the other object;
68
+ - **look swap**: the two objects exchange colour or size, so the same words now name the other physical object;
69
+ - **both**: both swaps at once.
70
+
71
+ Every scene has exactly one object that matches its words (checked by a label audit, column `audit`). The planned
72
+ positions are written straight into the simulator start state, so all scenes of a family share every pose. Scenes
73
+ run in LIBERO on robosuite (MuJoCo).
74
+
75
+ ### Groups and splits
76
+
77
+ | Group | Split | Families | Scenes |
78
+ |---|---|---:|---:|
79
+ | conflict_and_absent_word | absent_word | 80 | 960 |
80
+ | conflict_and_absent_word | conflict_colour_category | 80 | 640 |
81
+ | conflict_and_absent_word | conflict_colour_size | 40 | 320 |
82
+ | conflict_and_absent_word | conflict_difficulty | 180 | 720 |
83
+ | conflict_and_absent_word | conflict_pattern_colour | 40 | 320 |
84
+ | conflict_and_absent_word | conflict_size_category | 80 | 640 |
85
+ | language | adjective_order_cs_topup | 60 | 480 |
86
+ | language | q1_negation | 60 | 240 |
87
+ | language | q1_negation_topup2 | 270 | 540 |
88
+ | language | q2_size_rank | 60 | 360 |
89
+ | language | q3_colour_forms | 50 | 250 |
90
+ | language | q3_size_forms | 50 | 250 |
91
+ | language | q4_book | 60 | 240 |
92
+ | language | q4_bowl | 60 | 240 |
93
+ | language | q4_mug | 60 | 240 |
94
+ | language | q8_three | 60 | 240 |
95
+ | language | q8_two | 60 | 240 |
96
+ | sweeps | sweep_clutter | 200 | 800 |
97
+ | sweeps | sweep_clutter_topup | 60 | 240 |
98
+ | sweeps | sweep_hue | 150 | 600 |
99
+ | sweeps | sweep_hue_topup | 50 | 200 |
100
+ | sweeps | sweep_size | 150 | 600 |
101
+ | vla_bind_v1 | colour | 80 | 320 |
102
+ | vla_bind_v1 | conjunction | 120 | 960 |
103
+ | vla_bind_v1 | conjunction_matched | 60 | 480 |
104
+ | vla_bind_v1 | heldout_colour | 40 | 160 |
105
+ | vla_bind_v1 | nuisance | 40 | 280 |
106
+ | vla_bind_v1 | pattern_cup_block | 40 | 160 |
107
+ | vla_bind_v1 | size | 60 | 240 |
108
+ | vla_bind_v1 | size_transposition | 60 | 240 |
109
+
110
+ `vla_bind_v1` is VLA-Bind version 1.0: 2,840 scenes in eight splits. Six of them (colour, held-out colour, size,
111
+ size transposition, conjunction, nuisance) form the 2,200-scene clean set. `conjunction_matched` is the
112
+ matched-size ramekin split and `pattern_cup_block` the pattern split.
113
+
114
+ - **Sweeps** change one thing at a time. Hue moves a second bowl from another trained hue towards the named hue,
115
+ in 5 steps from 0 to 0.85 of the way. Size ratio runs in 4 steps from 1.5 down to 1.1. Clutter adds 0 to 3 extra
116
+ objects that share either the named colour or the named noun. The `_topup` splits add episodes to the same
117
+ sweeps.
118
+ - **Conflict scenes** put an object that matches one word against an object that matches a second word, for pairs
119
+ drawn from colour, size, kind and pattern. `absent_word` names an attribute that no object has.
120
+ - **Language sets** cover negation ("the bowl that is not red"), held-out forms (new words and sentence frames on
121
+ meanings seen in training), size rank, and adjective word order.
122
+
123
+ No seed or start state is shared between VLA-Bind and Train-Bind.
124
+
125
+ ### Scene table columns
126
+
127
+ | Column | Meaning |
128
+ |---|---|
129
+ | `group`, `split` | Where the scene belongs |
130
+ | `family_id`, `family_index`, `family_seed`, `family_tags` | The family the scene comes from |
131
+ | `role` | The scene's role in the family (base, word swap, look swap, both, or a set-specific role) |
132
+ | `instruction` | The instruction given to the policy |
133
+ | `target` | Name of the one object that matches the instruction |
134
+ | `label`, `corner`, `rect` | Position of the scene in the family's 2 by 2 design |
135
+ | `tags`, `factors_json` | Reporting factors, such as the word, layout or sweep step |
136
+ | `edit_json` | The edit applied to the family's shared start state (moved objects, swaps, seed) |
137
+ | `expression_json`, `spec_json` | Full scene specification: objects, colours, sizes, candidates |
138
+ | `state_id` | First 16 hex digits of SHA-256 over the float64 bytes of the start state |
139
+ | `init_state` | The simulator start state (flattened MuJoCo state vector) |
140
+ | `audit` | Label audit result; `unique` means exactly one object matches |
141
+
142
+ To check a start state: `hashlib.sha256(np.asarray(init_state, np.float64).tobytes()).hexdigest()[:16] == state_id`.
143
+
144
+ ### How a run is scored
145
+
146
+ - The episode starts from `init_state` with `instruction`. It stops at the **first sustained contact**, and its
147
+ label is the object touched. A run that makes no contact counts as wrong. The label scores which object the
148
+ policy chose, not whether the grasp succeeded.
149
+ - **Accuracy**: the share of episodes whose first contact is `target`.
150
+ - **Word edge**: within a family, the share of word-swap pairs that the policy gets right both times. A policy that
151
+ ignores the words cannot pass it.
152
+ - **Conflict scenes**: the share of picks on the object that matches one of the two words. 0.5 means no preference.
153
+ Each conflict scene also runs with the two words exchanged.
154
+ - **Word order**: the change in the share of picks when the two adjectives swap order.
155
+ - **Blind run**: run every scene again with an empty instruction. This shows what the policy does from the image
156
+ alone, and a trained score above it comes from the words.
157
+
158
+ Each set fixes its episodes per seed, so a new policy can run exactly the episodes of the paper.
159
+
160
+ ## Train-Bind
161
+
162
+ 4,860 scripted demonstrations in four strata. The trainer skips no-op frames, which leaves the 388,236 frames
163
+ counted in the paper. This release keeps every frame.
164
+
165
+ | Stratum | Files | Demonstrations | Frames |
166
+ |---|---:|---:|---:|
167
+ | colour | (uploading) | 1,512 | (uploading) |
168
+ | conjunction | (uploading) | 1,080 | (uploading) |
169
+ | conjunction_matched | (uploading) | 540 | (uploading) |
170
+ | pattern_cup_block | (uploading) | 648 | (uploading) |
171
+ | size | (uploading) | 1,080 | (uploading) |
172
+
173
+ The `pattern_cup_block` stratum replaces an earlier pattern stratum and is the one used by every model in the paper.
174
+
175
+ ### Frame columns (`train_bind/data/*.parquet`)
176
+
177
+ | Column | Meaning |
178
+ |---|---|
179
+ | `episode_id` | `<file>/<demo>`, the key into `episodes.parquet` |
180
+ | `stratum`, `frame_index`, `is_last` | Position of the frame |
181
+ | `instruction` | The instruction of the demonstration |
182
+ | `action` | The commanded action, in the LIBERO action convention |
183
+ | `sim_state` | The simulator state at this frame |
184
+ | `ee_states`, `gripper_states`, `joint_states` | Robot proprioception |
185
+ | `agentview_image`, `wrist_image` | Front and wrist camera images (PNG, lossless). Stored as rendered, without flipping |
186
+
187
+ The original files are LIBERO-style HDF5. Every array is kept in these columns. HDF5 attributes are in `attrs_json`
188
+ of `episodes.parquet`.
189
+
190
+ ### Rewritten instructions
191
+
192
+ The LANG rows train on the same demonstrations with one rewritten instruction each:
193
+
194
+ | Kind | Demonstrations |
195
+ |---|---:|
196
+ | original (kept) | 1,254 |
197
+ | synonym | 1,328 |
198
+ | paraphrase | 1,309 |
199
+ | word order | 561 |
200
+ | negation | 408 |
201
+
202
+ `train_bind/instruction_rewrites/lang_rewrites_v2.parquet` has one row per demonstration: `episode_id` (the key into
203
+ `episodes.parquet`), `kind` and the rewritten `instruction`. `lists.json` holds the sentence frames, synonyms and
204
+ negation frames used, and the held-out words and frames that the rewrites avoid.
205
+
206
+ ## Loading
207
+
208
+ ```python
209
+ from datasets import load_dataset
210
+ scenes = load_dataset("LALABAB/VLA-Bind", "vla_bind_v1", split="train")
211
+ frames = load_dataset("LALABAB/VLA-Bind", "train_bind", split="train", streaming=True)
212
+ ```
213
+
214
+ ## Licence
215
+
216
+ Data: CC BY 4.0. The scenes run on LIBERO and robosuite, both MIT licensed.
217
+
218
+ ## Limitations
219
+
220
+ - All scenes are simulated tabletop scenes with a single arm. The benchmark does not test real-world transfer.
221
+ - Attributes covered: colour, size, kind and pattern, plus the language sets. Other attributes need a new look-swap
222
+ rule.
223
+ - First-contact scoring measures choice, not task success.
224
+
225
+ ## Citation
226
+
227
+ Under review. A citation will be added after the review.