maverickrzw/VR-X-SFT-RL

VR-X: Visual Reasoning Benchmark for UniVR

🌐 Project Page  |  📄 Paper  |  💻 Code  |  🤗 Model (Planning)  |  🤗 Model (General)

---

Overview

VR-X is the first large-scale benchmark designed for diverse and heterogeneous visual reasoning. It is curated from 1.5M raw samples across 16 diverse sources, rigorously filtered into:

Split	Samples	Purpose
SFT (Cold Initialization)	~310k	Supervised fine-tuning to instill visual reasoning priors
RL	~3k	Hard samples for VR-GRPO reinforcement learning
Benchmark Evaluation	1.8k	Held-out test set (not included in this release)

All samples follow a unified format: query image + textual instruction + visual reasoning trajectory, enabling models to learn diverse visual reasoning capabilities from pure visual demonstrations.

---

Task Categories

VR-X spans 6 task categories covering both long-horizon planning and general visual reasoning:

Category	Data Sources	Frames	Ratio	Description
Visual Guidance	EgoDex, Action100M, Epic-Kitchen, VideoCraftBench	454k	32.2%	Cooking, handcrafting, daily activities from first/third-person video
Robot Manipulation	AgiBot, Droid, Bridge, ZebraCoT-Robot	509k	30.6%	Robotic grasping, tool use, multi-step manipulation
Editing	ZebraCoT-Multiobject	128k	6.5%	Object manipulation and scene editing
Spatial Perception	ThinkMorph-Navigation, ZebraCoT-Embodiment	128k	14.4%	Navigation, spatial reasoning, embodied perception
Visual Search	VisualCoT, ThinkMorph-Search	44k	7.2%	Object localization, attention-based search
Puzzle & Game	VRBench, Zebra-Jigsaw, ThinkMorph-VisPuzzle	70k	9.1%	Mazes, jigsaw puzzles, visual reasoning games

---

Dataset Structure

VR-X-SFT-RL/
├── asset/
│   └── VR-X.png                  # Benchmark overview figure
│
├── VR-X/                         # ══ SFT split (310k samples) ══
│   ├── Agibot/                   # task_327.parquet, task_351.parquet, ...
│   ├── Bridge/
│   ├── Droid/
│   ├── EgoDex/
│   ├── Epic_Kitchen/
│   ├── Action100/
│   ├── VideoCraftBench/
│   ├── Zebra_Count/
│   ├── Zebra_Robot/
│   ├── Zebra_Jigsaw/
│   ├── Visual_Search/
│   ├── Visual_CoT_GQA/
│   ├── Spatial_Navigation/
│   ├── Spatial_Navigation_maze/
│   └── Spatial_Navigation_trapfield/
│
└── VR-X-RL/                      # ══ RL split (3k hard samples) ══
    ├── Agibot/
    ├── EgoDex/
    ├── VideoCraftBench/
    ├── Epic_Kitchen/
    ├── Zebra_Count/
    ├── Zebra_Jigsaw/
    └── Visual_Search/

Each subdirectory contains per-task .parquet files. Each row in a parquet file is one training sample.

---

Parquet Schema

Column	Type	Description
question	str	Textual instruction (e.g., "Tie the red rope around the white gift box. Finish this task in 3 steps.")
problem_image_bytes	list[bytes]	JPEG-encoded query/input frames (typically 1 frame)
answer_image_bytes	list[bytes]	JPEG-encoded answer/output frames (the visual reasoning trajectory)
problem_images	list[dict]	Per-frame metadata with a "caption" field for each query frame
answer_images	list[dict]	Per-frame metadata with a "caption" field for each answer frame
height	int	VQ token grid height (e.g., 32 → 512 px at 16× downsampling)
width	int	VQ token grid width (e.g., 40 → 640 px)
global_summary	str	Optional text summary of the full trajectory

---

Download

# Full dataset (SFT + RL splits)
huggingface-cli download maverickrzw/VR-X-SFT-RL --repo-type dataset --local-dir datasets/VR-X-SFT-RL

# Then symlink or move into the UniVR repo structure:
ln -s $(pwd)/datasets/VR-X-SFT-RL/VR-X     UniVR/datasets/VR-X
ln -s $(pwd)/datasets/VR-X-SFT-RL/VR-X-RL  UniVR/datasets/VR-X-RL

Or load directly in Python:

import pandas as pd
from pathlib import Path

# Load one source
df = pd.concat([
    pd.read_parquet(p)
    for p in Path("datasets/VR-X-SFT-RL/VR-X/Agibot").glob("*.parquet")
])

print(f"Samples: {len(df)}")
print(f"Columns: {list(df.columns)}")
print(f"Example instruction: {df.iloc[0]['question']}")

---

Usage with UniVR

SFT Training

Data sources and sampling caps are configured in UniVR_SFT/train.py:

dataset_cfg = {
    "Agibot":                       {"max_samples_per_task": 300, "enabled": True},
    "Bridge":                       {"max_samples": 5000, "enabled": True},
    "Droid":                        {"max_samples": 5000, "enabled": True},
    "EgoDex":                       {"max_samples_per_task": 350, "enabled": True},
    "Epic_Kitchen":                 {"max_samples": 5000, "enabled": True},
    "Spatial_Navigation":           {"max_samples": 5000, "enabled": True},
    "Spatial_Navigation_maze":      {"max_samples": 5000, "enabled": True},
    "Spatial_Navigation_trapfield": {"max_samples": 5000, "enabled": True},
    "VideoCraftBench":              {"max_samples": 300, "repeat": 35, "enabled": True},
    "Visual_CoT_GQA":               {"max_samples": 10000, "enabled": True},
    "Visual_Search":                {"max_samples": 5000, "enabled": True},
    "Zebra_Count":                  {"max_samples": 10000, "enabled": True},
    "Zebra_Jigsaw":                 {"max_samples": 5000, "enabled": True},
    "Zebra_Robot":                  {"max_samples": 5000, "enabled": True},
    "Action100M":                   {"max_samples": 5000, "enabled": True},
}

RL Training

The RL split uses curated hard samples. Configured in UniVR_RL/examples/config_emu3.yaml:

data:
  train_files:
    agibot: ../datasets/VR-X-RL/Agibot
    egodex: ../datasets/VR-X-RL/EgoDex
    videocraft: ../datasets/VR-X-RL/VideoCraftBench
    epic_kitchen: ../datasets/VR-X-RL/Epic_Kitchen
    zebra_count: ../datasets/VR-X-RL/Zebra_Count
    zebra_jigsaw: ../datasets/VR-X-RL/Zebra_Jigsaw
    visual_search: ../datasets/VR-X-RL/Visual_Search

  source_sample_counts:
    agibot: 500
    egodex: 300
    epic_kitchen: 103
    videocraft: 51
    zebra_count: 500
    zebra_jigsaw: 200
    visual_search: 200

---

Data Curation Pipeline

VR-X is constructed via a four-stage pipeline:

1. Temporal Sampling: Raw video sequences are sampled using scene-aware detection (PySceneDetect) at 0.27 FPS, preserving richer information than random sampling.
2. QA Synthesis: Qwen3.5-397B synthesizes reasoning-oriented questions and textual answers, with ~10 key steps per trajectory.
3. Key-frame Selection: Conditioned on QA pairs, the VLM identifies the most relevant query image and key-step frames.
4. Quality Filtering: Rigorous filtering removes ~80% of candidates, discarding image-text mismatches, trivial questions, and low visual quality samples.

For non-video sources (VisualCoT, ZebraCoT), data is standardized directly from existing image sequences.

---

Citation

@article{ren2026univr,
  title={UniVR: Thinking in Visual Space for Unified Visual Reasoning},
  author={Zhongwei Ren and Yunchao Wei and Zhao Yao and Guixun Luo and Yao Zhao and Weibo Gong and Xiao Liu and Anran Wang and Xiangtai Li and Xiaojie Jin},
  year={2026},
}

License

This dataset is released under the Apache 2.0 License.