VLA-Bind / README.md
LALABAB's picture
Card: note on ids inside files
708db62 verified
|
Raw History Blame Contribute Delete
11.4 kB
---
license: cc-by-4.0
pretty_name: VLA-Bind and Train-Bind
language:
- en
task_categories:
- robotics
tags:
- vision-language-action
- attribute-binding
- benchmark
- libero
- robosuite
- manipulation
size_categories:
- 100K<n<1M
configs:
- config_name: vla_bind_v1
data_files: benchmark/scenes/vla_bind_v1.parquet
default: true
- config_name: sweeps
data_files: benchmark/scenes/sweeps.parquet
- config_name: conflict_and_absent_word
data_files: benchmark/scenes/conflict_and_absent_word.parquet
- config_name: language
data_files: benchmark/scenes/language.parquet
- config_name: train_bind
data_files: train_bind/data/*.parquet
- config_name: train_bind_episodes
data_files: train_bind/episodes.parquet
---
# VLA-Bind and Train-Bind
VLA-Bind is a benchmark for attribute binding in vision-language-action (VLA) policies. It asks one question: when
an instruction names an object by an attribute ("the red bowl", "the smaller cup"), does the policy pick the object
the words name, even when other objects share part of the description? Train-Bind is the matching training set of
scripted demonstrations.
This repository holds the data of the paper *Lost Among Look-Alikes: A Benchmark for Attribute Binding in
Vision-Language-Action Policies* (under review). The trained adapters are in
[LALABAB/VLA-Bind-adapters](https://huggingface.co/LALABAB/VLA-Bind-adapters). The scene generator and the evaluation code are not
part of this release yet. This card describes the data well enough to load and score it without them.
All results in the paper use **version 1.0**. Fixes will get a new version number, and version 1.0 stays available.
## What is in the release
| Folder | What it holds |
|---|---|
| `benchmark/scenes/` | One parquet table per group, one row per scene, with its start state |
| `benchmark/manifests/` | The same scenes in their original form: one JSON file per split and one `.npz` of start states |
| `train_bind/data/` | Train-Bind demonstrations, one row per frame, with both camera images |
| `train_bind/episodes.parquet` | One row per demonstration: instruction, stratum, start state, generation attributes |
| `train_bind/meta/` | Generation settings per file and the balance report |
| `train_bind/subsets/` | The 720 demonstrations used to train the `train_bind_720_demos` adapters |
| `train_bind/instruction_rewrites/` | The rewritten instruction of every demonstration used by the `rewritten_instructions` adapters, and the word lists behind them |
## The benchmark
### Scene families
A family is a set of scenes built from one table state, four in the basic case:
- **base**: the words name one object;
- **word swap**: the words change so that they name the other object;
- **look swap**: the two objects exchange colour or size, so the same words now name the other physical object;
- **both**: both swaps at once.
Every scene has exactly one object that matches its words (checked by a label audit, column `audit`). The planned
positions are written straight into the simulator start state, so all scenes of a family share every pose. Scenes
run in LIBERO on robosuite (MuJoCo).
### Groups and splits
| Group | Split | Families | Scenes |
|---|---|---:|---:|
| conflict_and_absent_word | absent_word | 80 | 960 |
| conflict_and_absent_word | conflict_colour_category | 80 | 640 |
| conflict_and_absent_word | conflict_colour_size | 40 | 320 |
| conflict_and_absent_word | conflict_difficulty | 180 | 720 |
| conflict_and_absent_word | conflict_pattern_colour | 40 | 320 |
| conflict_and_absent_word | conflict_size_category | 80 | 640 |
| language | adjective_order_colour_size | 60 | 480 |
| language | negation | 60 | 240 |
| language | negation_extra | 270 | 540 |
| language | size_rank | 60 | 360 |
| language | colour_held_out_forms | 50 | 250 |
| language | size_held_out_forms | 50 | 250 |
| language | new_category_book | 60 | 240 |
| language | new_category_bowl_control | 60 | 240 |
| language | new_category_mug | 60 | 240 |
| language | no_attribute_word_three_bowls | 60 | 240 |
| language | no_attribute_word_two_bowls | 60 | 240 |
| sweeps | sweep_clutter | 200 | 800 |
| sweeps | sweep_clutter_extra | 60 | 240 |
| sweeps | sweep_hue | 150 | 600 |
| sweeps | sweep_hue_extra | 50 | 200 |
| sweeps | sweep_size | 150 | 600 |
| vla_bind_v1 | colour | 80 | 320 |
| vla_bind_v1 | conjunction | 120 | 960 |
| vla_bind_v1 | conjunction_matched_size | 60 | 480 |
| vla_bind_v1 | heldout_colour | 40 | 160 |
| vla_bind_v1 | nuisance | 40 | 280 |
| vla_bind_v1 | pattern | 40 | 160 |
| vla_bind_v1 | size | 60 | 240 |
| vla_bind_v1 | size_transposition | 60 | 240 |
`vla_bind_v1` is VLA-Bind version 1.0: 2,840 scenes in eight splits. Six of them (colour, held-out colour, size,
size transposition, conjunction, nuisance) form the 2,200-scene clean set. `conjunction_matched_size` is the
matched-size ramekin split and `pattern` the pattern split.
- **Sweeps** change one thing at a time. Hue moves a second bowl from another trained hue towards the named hue,
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
objects that share either the named colour or the named noun. The `_extra` splits add episodes to the same
sweeps.
- **Conflict scenes** put an object that matches one word against an object that matches a second word, for pairs
drawn from colour, size, kind and pattern. `absent_word` names an attribute that no object has.
- **Language sets** cover negation ("the bowl that is not red"), held-out forms (new words and sentence frames on
meanings seen in training), size rank, and adjective word order.
The word-order sets with trained and fresh frames, the fresh negation frames, word shuffle, tie and the three-object sets were also evaluated in the paper; they will be added in a later upload. The `split` column of the scene tables and the ids inside the manifests still use the earlier split ids; the same upload will bring them in line with the names above.
No seed or start state is shared between VLA-Bind and Train-Bind.
### Scene table columns
| Column | Meaning |
|---|---|
| `group`, `split` | Where the scene belongs |
| `family_id`, `family_index`, `family_seed`, `family_tags` | The family the scene comes from |
| `role` | The scene's role in the family (base, word swap, look swap, both, or a set-specific role) |
| `instruction` | The instruction given to the policy |
| `target` | Name of the one object that matches the instruction |
| `label`, `corner`, `rect` | Position of the scene in the family's 2 by 2 design |
| `tags`, `factors_json` | Reporting factors, such as the word, layout or sweep step |
| `edit_json` | The edit applied to the family's shared start state (moved objects, swaps, seed) |
| `expression_json`, `spec_json` | Full scene specification: objects, colours, sizes, candidates |
| `state_id` | First 16 hex digits of SHA-256 over the float64 bytes of the start state |
| `init_state` | The simulator start state (flattened MuJoCo state vector) |
| `audit` | Label audit result; `unique` means exactly one object matches |
To check a start state: `hashlib.sha256(np.asarray(init_state, np.float64).tobytes()).hexdigest()[:16] == state_id`.
### How a run is scored
- The episode starts from `init_state` with `instruction`. It stops at the **first sustained contact**, and its
label is the object touched. A run that makes no contact counts as wrong. The label scores which object the
policy chose, not whether the grasp succeeded.
- **Accuracy**: the share of episodes whose first contact is `target`.
- **Word edge**: within a family, the share of word-swap pairs that the policy gets right both times. A policy that
ignores the words cannot pass it.
- **Conflict scenes**: the share of picks on the object that matches one of the two words. 0.5 means no preference.
Each conflict scene also runs with the two words exchanged.
- **Word order**: the change in the share of picks when the two adjectives swap order.
- **Blind run**: run every scene again with an empty instruction. This shows what the policy does from the image
alone, and a trained score above it comes from the words.
Each set fixes its episodes per seed, so a new policy can run exactly the episodes of the paper.
## Train-Bind
4,860 scripted demonstrations in four strata. The trainer skips no-op frames, which leaves the 388,236 frames
counted in the paper. This release keeps every frame.
| Stratum | Files | Demonstrations | Frames |
|---|---:|---:|---:|
| colour | 63 | 1,512 | 120,669 |
| conjunction | 45 | 1,080 | 85,042 |
| conjunction_matched_size | 23 | 540 | 43,019 |
| pattern | 27 | 648 | 53,061 |
| size | 45 | 1,080 | 86,446 |
The `pattern` stratum replaces an earlier pattern stratum and is the one used by every model in the paper.
### Frame columns (`train_bind/data/*.parquet`)
| Column | Meaning |
|---|---|
| `episode_id` | `<file>/<demo>`, the key into `episodes.parquet`. Ids inside the files keep the original file names, so rows of `colour_000.parquet` have ids `tb_colour_000/demo_<n>`, and the pattern stratum is stored as `pattern_cup_block` |
| `stratum`, `frame_index`, `is_last` | Position of the frame |
| `instruction` | The instruction of the demonstration |
| `action` | The commanded action, in the LIBERO action convention |
| `sim_state` | The simulator state at this frame |
| `ee_states`, `gripper_states`, `joint_states` | Robot proprioception |
| `agentview_image`, `wrist_image` | Front and wrist camera images (PNG, lossless). Stored as rendered, without flipping |
The original files are LIBERO-style HDF5. Every array is kept in these columns. HDF5 attributes are in `attrs_json`
of `episodes.parquet`.
### The 720-demonstration subset
`train_bind/subsets/train_bind_720_demos_subset.json` lists the 720 demonstrations (296 colour, 212 size, 212 conjunction) used to train the `train_bind_720_demos` adapters, with their seeds, state ids and family ids. Join on `family_id` (in `attrs_json` of `episodes.parquet`) to find their frames.
### Rewritten instructions
The `rewritten_instructions` adapters train on the same demonstrations with one rewritten instruction each:
| Kind | Demonstrations |
|---|---:|
| original (kept) | 1,254 |
| synonym | 1,328 |
| paraphrase | 1,309 |
| word order | 561 |
| negation | 408 |
`train_bind/instruction_rewrites/lang_rewrites_v2.parquet` has one row per demonstration: `episode_id` (the key into
`episodes.parquet`), `kind` and the rewritten `instruction`. `lists.json` holds the sentence frames, synonyms and
negation frames used, and the held-out words and frames that the rewrites avoid.
## Loading
```python
from datasets import load_dataset
scenes = load_dataset("LALABAB/VLA-Bind", "vla_bind_v1", split="train")
frames = load_dataset("LALABAB/VLA-Bind", "train_bind", split="train", streaming=True)
```
## Licence
Data: CC BY 4.0. The scenes run on LIBERO and robosuite, both MIT licensed.
## Limitations
- All scenes are simulated tabletop scenes with a single arm. The benchmark does not test real-world transfer.
- Attributes covered: colour, size, kind and pattern, plus the language sets. Other attributes need a new look-swap
rule.
- First-contact scoring measures choice, not task success.
## Citation
Under review. A citation will be added after the review.