RoboSteer

Benchmarking Behavioral Steerability in Behavior Foundation Models

3 levels · 20 task families · 773,716 task definitions

Paper · Project website · Explore examples · Download · Appendix B.3 figures

RoboSteer evaluates whether a behavior foundation model (BFM) can generate humanoid motion that follows a user's intent. The paper calls this ability behavioral steerability: a motion should satisfy the requested behavior and any conditions on how it is performed.

The benchmark has three levels:

Level	What the model must do	Example
1 · Conditional steering	Generate or complete a motion from a supplied condition	Follow a text description or finish a partial motion
2 · Constraint steering	Preserve the intended action while obeying an added requirement	Move faster or reach a target location
3 · Compositional steering	Follow multiple, ordered requirements from different input sources	Combine video, text, audio, and key-frame instructions in sequence

The release contains task-definition JSON files and the motion, video, audio, and image assets they reference. A task's input describes the conditions given to a model; ground_truth points to reference or target motion. The reference media shown here are not model predictions.

Start here

1. Explore without downloading: use the Dataset Viewer above. Choose level1_examples, level2_examples, or level3_examples to inspect 22 selected tasks across the three levels. The examples split is a browsing sample, not a training/test partition or the full benchmark.
2. See what each task looks like: browse the 48 Appendix B.3 case figures, organized by task and input modality. Each figure has an in-page image, a full-size PNG, and its original PDF. Media examples show actual input assets and task JSONs.
3. Read a real task: follow the worked JSON example, then use the task map and field reference.
4. Get the complete release: follow the download and extraction steps. The 23 archives total approximately 355.35 GiB.

In the viewer, a reference_motion_video thumbnail plays the source/reference motion. It is not a generated result. For tasks with several clips or timed Level 3 inputs, follow the ordering and timestamps in the task JSON; the media pages preserve those details.

1. Download the dataset

The complete release is hosted in this repository. It contains 23 independent TAR archives, preserving the original file bytes and paths.

• Original data: 352.05 GiB across 1,923,050 files.
• Archives: approximately 355.35 GiB. Keeping both archives and extracted data needs about 707.40 GiB, plus filesystem overhead and download cache space.
• Download all archive groups for the complete benchmark. Tasks can reference assets in Data/Shared/; a task's Level does not fully specify its download dependencies.

Install the Hugging Face client and download into RoboSteer-download:

python -m pip install --upgrade huggingface_hub
hf download YanCORANV/RoboSteer --repo-type dataset --include "archives/**" "SHA256SUMS" --local-dir RoboSteer-download

If interrupted, run the same download command with the same local directory to reuse completed downloads and available cache state.

2. Extract into one dataset root

RoboSteer-download holds the downloaded TAR files. Create a separate destination, Steerable Motion Benchmark Dataset, for the extracted data. This destination can be anywhere on your disk; retain its internal folder structure.

Run the following in PowerShell, from the parent of RoboSteer-download. It checks each archive and extracts it into the same root. Keep the archives until extraction succeeds.

PowerShell: verify and extract all archives (resumable)

$download = (Resolve-Path -LiteralPath './RoboSteer-download').Path
$dataset = Join-Path (Get-Location).Path 'Steerable Motion Benchmark Dataset'
New-Item -ItemType Directory -Path $dataset -Force | Out-Null
$completedFile = Join-Path $download '.extracted-sha256.txt'
$completed = @()
if (Test-Path -LiteralPath $completedFile) {
    $completed = @(Get-Content -LiteralPath $completedFile)
}
$checks = @(Get-Content -LiteralPath (Join-Path $download 'SHA256SUMS') | Where-Object { $_.Trim() })
$index = 0
foreach ($line in $checks) {
    $index++
    if ($line -notmatch '^([0-9a-fA-F]{64})\s+\*?(.+)$') { throw "Invalid checksum line: $line" }
    $expected = $Matches[1].ToLowerInvariant()
    $relative = $Matches[2]
    $archive = Join-Path $download $relative
    $key = "$dataset|$expected|$relative"
    if ($completed -contains $key) { Write-Host "[$index/$($checks.Count)] Already extracted: $relative"; continue }
    Write-Host "[$index/$($checks.Count)] Verify and extract: $relative"
    $actual = (Get-FileHash -LiteralPath $archive -Algorithm SHA256).Hash.ToLowerInvariant()
    if ($actual -ne $expected) { throw "Checksum mismatch: $relative" }
    & tar -xf $archive -C $dataset
    if ($LASTEXITCODE -ne 0) { throw "Extraction failed: $relative" }
    Add-Content -LiteralPath $completedFile -Value $key
}
Write-Host "Dataset root: $dataset"

The completion record lets you rerun this block after interruption. Do not reuse it if you delete or modify the extracted files. A partial archive is extracted again on restart.

Your local layout should be:

Steerable Motion Benchmark Dataset/
├── Data/
│   ├── Level1/    Audio, Image, Motion, Spatial, Video
│   ├── Level2/    Audio, Order, Times, Trajectory
│   ├── Level3/    Audio, Image, Video
│   └── Shared/    Metadata, Motion, Video/Human, Video/Skeleton
└── Tasks/
    ├── Level1/
    │   ├── full_conditioning_reproduction/
    │   ├── spatial_completion/
    │   └── temporal_completion/
    ├── Level2/    Amplitude, BodyRestrain, Direction, Order, Speed, Times, Trajectory
    └── Level3/    task JSON files

There must be one Data/ and one Tasks/ directly under the dataset root. Do not extract each TAR into its own folder. Paths inside task JSONs are relative to this root, not to the JSON's containing folder.

3. Read your first task

Tasks/ describes what to do; Data/ contains the assets those tasks reference. The following complete example opens a real text-to-motion task, prints its condition, then reads the first frame of its referenced joint positions. It requires only Python's standard library.

Save as read_first_task.py, run python read_first_task.py, and enter your extracted dataset root when prompted. A ready-to-save copy is available here.

import csv
import json
from pathlib import Path

root = Path(input("Extracted dataset root: ").strip().strip('"')).expanduser().resolve()
task_path = root / 'Tasks/Level1/full_conditioning_reproduction/text_to_motion_generation/L1_text_to_motion_generation_text_--04DHOI32k_00009_38_121.json'
task = json.loads(task_path.read_text(encoding="utf-8-sig"))
print("Task:", task["metadata"]["task_id"])
print("Instruction:", task["input"]["prompts"]["general_instruction"])
print("Text condition:", task["input"]["modalities"]["text"])
motion = root / task["ground_truth"]["motion_parameters"]
with (motion / "joint_pos.csv").open(encoding="utf-8-sig", newline="") as f:
    reader = csv.reader(f)
    columns = next(reader)
    first_frame = [float(value) for value in next(reader)]
print("Motion directory:", motion)
print("Joint columns:", len(columns))
print("First frame:", first_frame)

For this sample, the printed joint-column count is 29. You have now read an actual task and its motion reference. This example does not run a model or evaluate predictions.

Keep input separate from ground_truth: references are not additional conditioning inputs. For Level 3, read the ordered input.interleave.sequence instead of expecting ordinary modality lists.

4. Choose a benchmark task

Level	Organization	Task definitions
Level 1	Full conditioning, temporal completion, spatial completion	630,623
Level 2	Amplitude, Body Restrain, Direction, Order, Speed, Times, Trajectory	135,329
Level 3	Ordered multimodal interleaving	7,764

Use the task and folder map for every task-bearing directory, counts, input types, real JSON examples, and exact asset-path examples. Several tasks share motion assets; these counts are task definitions, not unique motion sequences.

5. Understand the task JSON

Section	How to use it
metadata	Identify the level/family, duration, and source references
input.prompts	Read the general instruction and any modifier
input.modalities	Load text, audio, video, image, or spatial conditions for Levels 1/2
input.interleave.sequence	Preserve the ordered, timed components for Level 3
ground_truth	Locate the task's target/reference motion and any temporal or trajectory constraints

The field-by-field JSON reference explains nested fields, types, units, empty values, image representation differences, and task-specific reference semantics. In particular, the Level 2 Amplitude/Speed/Direction/Body Restrain references are unmodified source motions, not precomputed modified outputs.

Motion packages are not all the same format: shared G1 motion uses CSV directories, while Order/Times use AMASS/BABEL-linked PKL files. Follow the actual task path and the format reference.

Benchmark integration

The dataset root is configurable in your own reader. See the project website for the paper and code resources. This card documents the dataset layout and one reading example; use the released evaluation instructions for model adapters and scoring. The saved prompt text is preserved as released.

Release details

The full upload completed on 2026-09-17. All 23 archive sizes and checksums were verified against the local upload records. See SHA256SUMS and UPLOAD_STATUS.json. The example gallery is an additional browsing layer; it does not replace or alter the original archives.

Data sources and licensing status

The contributors identify Level 2 Order and Times as using AMASS motion data and BABEL annotations, and identify the other task data as their own. The full-directory transfer includes the existing Order and Times materials; it is not a third-party-data-excluded release.

The repository does not assign a blanket license to all contents. The license for the team's own data and the redistribution scope for third-party-derived materials remain under review. Repository visibility and download availability are not a grant of redistribution or commercial-use rights.

• AMASS project · AMASS official license
• BABEL project · BABEL official license

The official licenses include non-commercial-use and no-distribution provisions. Refer to the applicable source agreements and any additional written permissions. No additional redistribution authorization is asserted here.

Citation

Minghe Gao, Zhanxi Yan, Jiahui Liu, Wendong Bu, Xiaoting Chen, Qizhou Wang, Yi Su, Siliang Tang, Jun Xiao, Yueting Zhuang, Tat-Seng Chua, and Juncheng Li. Benchmarking Behavioral Steerability in Behavior Foundation Models, arXiv:2610.10198, 2026.

@misc{gao2026benchmarking,
  title={Benchmarking Behavioral Steerability in Behavior Foundation Models},
  author={Minghe Gao and Zhanxi Yan and Jiahui Liu and Wendong Bu and Xiaoting Chen and Qizhou Wang and Yi Su and Siliang Tang and Jun Xiao and Yueting Zhuang and Tat-Seng Chua and Juncheng Li},
  year={2026},
  eprint={2610.10198},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.10198}
}

For AMASS and BABEL, also use the citations supplied by their official project pages where applicable.