File size: 7,106 Bytes
c323605 | 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 | ---
license: mit
pretty_name: BEHAVIOR-Spatial
task_categories:
- visual-question-answering
- robotics
tags:
- spatial-reasoning
- embodied-ai
- behavior-1k
- lerobot
- arxiv:2610.34051
---
# BEHAVIOR-Spatial Benchmark
Data for **BEHAVIOR-Spatial Benchmark: Evaluating Spatial Understanding for Robotics**
([arXiv:2610.34051](https://arxiv.org/abs/2610.34051)). The benchmark measures whether spatial understanding in
vision-language models (VLMs) carries over to vision-language-action (VLA)
behavior. It has three components, all built in BEHAVIOR-1K / OmniGibson:
| Component | What | Size |
|---|---|---|
| **BEHAVIOR-Spatial-VQA** (Understand) | multiple-choice spatial questions | 3,256 questions: 1,572 general + 1,082 manipulation-task + 602 navigation-task |
| **BEHAVIOR-Spatial-IR** (Ground) | point-based target grounding: predict the instructed target's image point | 1,240 queries over 14 manipulation tasks |
| **BEHAVIOR-Spatial-Action** (Act) | sim evaluation instances and LeRobot training sets | 14 manipulation + 10 navigation tasks |
Project website: https://arielchen07.github.io/behavior-spatial/
Code: https://github.com/arielchen07/BEHAVIOR-Spatial-benchmark
## Layout
```
vqa/
general/single_frame/ 1,078 questions (direction 600 / distance 339 / size 139), images/
general/multi_frame/rotation/ 239 questions (direction / distance / size)
general/multi_frame/navigation/ 255 questions
task/manipulation/<task>/ 1,082 task-grounded questions over 14 manipulation tasks
task/navigation/<task>/ 602 task-grounded questions over 10 navigation tasks (+ context/ frames)
ir/
gt/<task>.json 1,240 grounding queries (14 tasks)
frames/<task>/ 720x720 head-camera frames
masks/<task>/ instance-segmentation masks (12 tasks; see "IR ground truth")
action/
eval/manipulation/held_out/ 469 evaluation instances (BDDL + task-relevant-object JSON)
eval/manipulation/prompt_probe/ 700 instruction-change probes
eval/navigation/held_out/ 401 navigation instances
train/manipulation/ LeRobot v2.1, 5,532 episodes (4,833 unique; 14 tasks + training-only place_saucer)
train/navigation/ LeRobot v2.1, 3,912 episodes
train/manipulation_subsets/ episode-index lists for the 30% / 50% data-scaling runs
train/navigation_context/ context-tour frames per scene (21 / 6 / 13 / 16 / 21)
scripts/downscale_to_224.py rebuild the 224x224 inputs the paper's policies trained on
MANIFEST.json per-component counts + sha256 of every JSON/JSONL file
```
Each VQA `questions.json` holds records with `question_id`, `question`,
`options`, `gt_answer`, and image path(s) relative to that file's directory.
## Training data (LeRobot v2.1)
`action/train/{manipulation,navigation}` load with stock LeRobot v2.1
(`LeRobotDataset(repo_id, root=...)`) and use the standard
`chunk-{episode_index // 1000:03d}` layout.
State and action are 23-dim (base velocity 3, trunk 4, left arm 7, left gripper 1,
right arm 7, right gripper 1) at 30 fps. Normalization statistics used in the paper
are in `norm_stats.json`; per-task sampling weights in `task_weights.json`.
All video streams are at native resolution:
| key | manipulation | navigation | resolution |
|---|---|---|---|
| `head_image` | ✓ | ✓ | 720×720 |
| `left_wrist_image`, `right_wrist_image` | ✓ | | 480×480 |
| `object_center_image` (dot IR) | ✓ | | 720×720 |
| `object_center_first_image` (frame-0 dot IR, repeated) | ✓ | | 720×720 |
| `attn_map_image` (Gaussian attention-map IR) | ✓ | | 720×720 |
- **Resolution used in the paper.** The paper's policies were trained on a
224×224 downscale of these streams; `scripts/downscale_to_224.py` rebuilds it.
- **`meta/ir_centers.jsonl`** (manipulation) gives the per-frame target center in
720×720 head-image pixels (`null` = out of view; placement tasks use the
destination). It is all you need to render the IR at any resolution.
- **Navigation is head-camera only**, matching the paper's navigation policies.
- **Manipulation training covers 15 tasks**: the 14 benchmark tasks plus
`place_saucer` (97 episodes), which is used only for training, as in the paper.
- **Repeated episodes.** The 5,532 manipulation episodes are 4,833 unique
demonstrations plus 699 exact copies, kept so the set matches what the paper's
policies trained on: variant balancing for `reach_food_item` (416) and
`place_bowl_in_cabinet` (89), and upweighting of the approach-augmented
`reach_paperback_book` episodes (194). `meta/duplicate_of.json` maps each copy to
its original; drop its keys for unique demonstrations. `task_weights.json` adds
per-task sampling weights on top (default preset `uniform_by_task`).
Sources: BEHAVIOR-1K 2025 and 2026 challenge demonstrations
(`behavior-1k/2025-challenge-demos`, `behavior-1k/2026-challenge-demos`) plus
simulator replays for `reach_hardback_from_bookcase`, part of `reach_plate`, and
the approach-augmented `reach_paperback_book` episodes.
## IR ground truth
`ir/gt/<task>.json` records carry `instruction`, `gt_target`, `candidates`, `image_path`
and per-candidate `bboxes_2d` (720×720 pixels), plus the camera and scene geometry the
scorer uses. Placement tasks are scored on the destination; tolerant and strict accuracy
are defined in the paper.
12 tasks also have a `mask_path`: a single-channel 720×720 PNG of instance labels
(pixel value `i` = `mask_names[i-1]`, `0` = background; it looks black in an image
viewer). A predicted point scores on the object whose mask it lands on.
`reach_paperback_book` and `reach_hardback_from_bookcase` are scored on the boxes.
## Evaluation instances
`action/eval/**` holds the BDDL task definitions and task-relevant-object (`_tro.json`)
files that the simulator harness loads. The 2025-scene tasks run on OmniGibson 3.7.2;
the four restaurant-diner navigation tasks need OmniGibson 3.9.2 / Isaac Sim 5.1
(BEHAVIOR-1K 2026 assets). Each instance's `goal_conditions` sets how it is scored.
The VQA prompt builder, the VQA / IR scorers and the simulator harness are in the
[code release](https://github.com/arielchen07/BEHAVIOR-Spatial-benchmark).
## License
This dataset is released under the MIT License, matching the BEHAVIOR-1K demonstrations
it builds on. It contains rendered images and videos, annotations, and task definitions,
but no 3D assets. Running the simulator evaluation requires the BEHAVIOR-1K scene and
object assets, which are distributed separately under the
[BEHAVIOR Data Bundle End User License Agreement](https://behavior.stanford.edu/)
(non-commercial academic research only).
## Citation
```bibtex
@article{chen2026behaviorspatial,
title = {{BEHAVIOR-Spatial} Benchmark: Evaluating Spatial Understanding for Robotics},
author = {Chen, Sirui and Huang, Huang and Wang, Qineng and Huang, Wenlong and Li, Manling and Zhang, Ruohan and Fei-Fei, Li and Wu, Jiajun},
journal = {arXiv preprint arXiv:2610.34051},
year = {2026}
}
```
|