---
license: cc-by-nc-4.0
pretty_name: EmbRACE
language:
- en
tags:
- embodied-ai
- vision-language-navigation
- egocentric
- unreal-engine
- simulation
- benchmark
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: dataset
    path: data/dataset.parquet
  - split: benchmark
    path: data/benchmark.parquet
---

# EmbRACE: Embodied Reasoning and Action in Complex Environments

[Viewer](https://huggingface.co/spaces/mxlin043/EmbRACE-Viewer) · [Paper](https://arxiv.org/abs/2507.10548) · [Code](https://github.com/mxlin043/EmbRACE)

EmbRACE consists of two parts. The dataset is a set of 3,421 human demonstrations for closed-loop tasks in 55 environments, 48,264 steps in all, in which every step is paired with a rationale written from the agent's viewpoint and verified by annotators. The benchmark is a set of 686 tasks in 7 further environments, run in the closed loop and scored on the agent's final position and on the door and object states.

| Part | Environments | Tasks | Steps | Rationales |
|---|---:|---:|---:|---:|
| dataset | 55 | 3,421 | 48,264 | 48,264 |
| benchmark | 7 | 686 | 10,552 | 0 |

| Type | Task type | dataset | benchmark |
|---:|---|---:|---:|
| 0 | Basic | 646 | 116 |
| 1 | Exploration | 591 | 117 |
| 2 | Dynamic Spatial-Semantic | 686 | 114 |
| 3 | Multi-stage | 644 | 106 |
| 4 | Open Door | 226 | 109 |
| 5 | Pick & Drop | 628 | 124 |

- **Basic**: The target is visible from the start pose and directly reachable.
- **Exploration**: The target is out of view at the start, so the agent searches the environment and commits to a direction before the target comes into view.
- **Dynamic Spatial-Semantic**: The target is visible but specified relationally or ordinally, for example "the second chair from the left". The relation is defined in the initial view.
- **Multi-stage**: The task names two targets to be reached in order.
- **Open Door**: The agent opens a door. Success is judged on the door state.
- **Pick & Drop**: The agent picks up an object and releases it beside a landmark. Success is judged on the object position.

## Files

```
dataset/task/<Env>/<Task>.json            task files
dataset/trajectory/<Env>/<Task>/          frames and record of each demonstration
benchmark/task/<Env>/<Task>.json          task files read by the evaluation code
benchmark/trajectory/<Env>/<Task>/        reference human demonstration of each task
benchmark/success_region/<Env>/<Task>/    success regions of each task
packed/{dataset,benchmark}/<Env>.tar      all files of one environment in one archive
data/{dataset,benchmark}.parquet          one row per task with its initial frame
index/web/tasks.json                      task index read by the viewer
environments.json                         the 62 environments
pd_objects.json                           the ten Pick & Drop objects
SHA256SUMS                                checksums of all files
```

This is the layout that the archives in `packed/` hold and that the evaluation code reads. The files shown on this page are a copy for browsing, in which every task folder and task file carries the prefix `task_`, for example `benchmark/trajectory/CourtYard/task_x001312_y007194_z000118_t0_a/`, because the Hub does not accept some of the original names. `SHA256SUMS` lists every file under its name in the archives.

## Download

The 62 archives hold every file under `dataset/` and `benchmark/`. Extracting them in one folder gives the layout above.

```bash
hf download mxlin043/EmbRACE --repo-type dataset --local-dir EmbRACE \
  --include "packed/*" --include "environments.json" --include "pd_objects.json"
cd EmbRACE && for f in packed/*/*.tar; do tar -xf "$f"; done
```

One environment can also be downloaded on its own, for example

```bash
hf download mxlin043/EmbRACE --repo-type dataset --local-dir EmbRACE --include "packed/benchmark/CourtYard.tar"
```

The task table loads with the `datasets` library.

```python
from datasets import load_dataset

tasks = load_dataset("mxlin043/EmbRACE", split="dataset")
```

Running the benchmark requires the UnrealZoo UE5.6 package of version v3.0.2 and the evaluation code.

## Format

### Task files

| Field | Content |
|---|---|
| `Instruction` | the instruction |
| `Type` | the task type, 0 to 5 |
| `Start_Pose` | position and orientation of the agent at the start, `[x, y, z, pitch, yaw, roll]` |
| `PD_ID`, `PD_Size` | Pick & Drop only, the object and its scale |
| `PD_Type` | Pick & Drop only, `pd` when the object is released beside a landmark and `p` when the task ends with the object held |
| `PD_Start_Pose` | Pick & Drop only, the pose at which the object is placed |
| `Exposure` | optional, an exposure offset for rendering that replaces the default of the environment |

A task is named by its start position in centimeters, its type and a letter, for example `x001312_y007194_z000118_t0_a`. The names of Pick & Drop tasks end in `_p` or `_pd` following `PD_Type`.

### Trajectory records

Each demonstration is one folder with its frames `000.jpg`, `001.jpg`, … at 640×480 and the record `infos.json`. The frames are the egocentric frame before every action and the frame after the last one, so a demonstration of n actions has n+1 frames. The pose and state streams have n+1 entries aligned with the frames. Poses are given in the coordinate frame of the environment, in centimeters and degrees.

| Field | Content |
|---|---|
| `Instruction`, `Type`, `Start_Pose`, `PD_ID`, `PD_Size`, `PD_Type`, `PD_Start_Pose` | as in the task file |
| `Action` | the demonstrated actions (n) |
| `Pose_Trajectory` | the agent pose (n+1) |
| `Door_State_Trajectory` | Open Door only, the door state, 0 closed and 1 open (n+1) |
| `PD_Pose_Trajectory` | Pick & Drop only, the object pose (n+1) |
| `PD_Dist_2D` | Pick & Drop only, the horizontal distance from the agent to the object (n+1) |
| `Rationale` | dataset only, the rationale of each step (n) |

The actions are `MoveForward`, `MoveBackward`, `TurnLeft`, `TurnRight`, `LookUp`, `LookDown`, `OpenDoor`, `Pick`, `Drop`, `MidwayTarget` and `Finish`. `MidwayTarget` declares arrival at the first target of a Multi-stage task.

### Success regions

Each benchmark task has an `area.json` with the regions it is scored against. Two images show the regions, `xy_region.png` as a top-down plot and `ue_top_down_overlay.jpg` drawn over the scene. Multi-stage tasks have one pair per target, prefixed `midway_` and `final_`.

| Field | Content |
|---|---|
| `coordinate_system` | the frame of the geometry, the XY plane of the environment in centimeters |
| `distance_metric` | distances are measured in the XY plane |
| `stages` | one entry per target, `final` and for Multi-stage tasks also `midway` |
| `character_regions` | the polygons in which the agent has to stand |
| `target_footprints` | the footprint of the target |
| `character_reference_z` | the reference height of the agent at the target |
| `pd_object_regions`, `pd_object_reference_z` | Pick & Drop only, the region and the reference height of the object |
| `shortest_path_uu` | the reference length of SPL, in centimeters |

### Other files

`environments.json` gives each environment its identifier, its name in the paper, its part and its setting (`id`, `name`, `split`, `setting`). `pd_objects.json` maps `PD_ID` to the name of the object. The Parquet tables in `data/` hold one row per task with its environment, type, instruction, initial frame, actions and rationales.

## License

The data are released under the [Creative Commons Attribution-NonCommercial 4.0](LICENSE) license, for research use with attribution. The environments are those of UnrealZoo and are obtained from its release under its terms.

## Citation

```bibtex
@article{lin2025embrace,
  title={EmbRACE: Embodied Reasoning and Action in Complex Environments},
  author={Lin, Mingxian and Huang, Wei and Li, Yitang and Jiang, Chengjie and Wu, Kui and Zhong, Fangwei and Chen, Weikai and Qian, Shengju and Wang, Xin and Qi, Xiaojuan},
  journal={arXiv preprint arXiv:2507.10548},
  year={2025}
}
```

## Changelog

This is version 1.0. Corrections to the data are recorded in `CHANGELOG.md`.