---
license: mit
task_categories:
- robotics
- visual-question-answering
- image-text-to-text
language:
- en
tags:
- scene-graph
- spatial-reasoning
- robot-manipulation
- vla
- vlm
- libero
- lerobot
pretty_name: EmbodimentSemantic
size_categories:
- 1K<n<10K
---

# EmbodimentSemantic

**A Spatial Scene-Graph Dataset and Benchmark for Vision-Language Models on Embodied Manipulation Trajectories**

Hassan Jaber¹ · Refinath S N² · Luca Cagliero¹ · Christopher E. Mower² · Haitham Bou-Ammar²³

¹Politecnico di Torino · ²Huawei Noah's Ark Lab · ³University College London

[![Paper](https://img.shields.io/badge/Paper-PDF-red?style=flat-square&logo=adobeacrobatreader&logoColor=white)](https://github.com/SemanticVLA/EmbodimentSemantic/blob/main/EmbodimentSemantic.pdf)
[![Code](https://img.shields.io/badge/Code-GitHub-black?style=flat-square&logo=github)](https://github.com/SemanticVLA/EmbodimentSemantic)

---

## Overview

EmbodimentSemantic is a benchmark dataset for evaluating whether vision-language models (VLMs) can recover exact spatial scene graphs from robot manipulation observations — and whether injecting those scene graphs into existing VLA policies improves downstream control.

The dataset has two components:

- **LIBERO Simulator Benchmark** — 500 demonstrations across 10 LIBERO-Spatial tasks (62,250 paired timesteps, 124,500 RGB frames). Ground-truth scene graphs are derived automatically from MuJoCo geometry, giving exact triplet-level supervision without manual annotation.
- **SO101 Real-Robot Dataset** — 257 teleoperated episodes across 5 tabletop bowl-placement tasks, collected with the low-cost SO101 arm. Includes external-camera, wrist-camera, and depth streams in LeRobot format.

---

## Files

| File | Size | Description |
|---|---|---|
| `libero_spatial_v5.zip` | 2.91 GB | LIBERO simulator benchmark: HDF5 demos with scene-graph annotations embedded under `obs/agentview_scene_graph` and `obs/robot0_eye_in_hand_scene_graph` |
| `SO1001_dataset.zip` | 6.13 GB | SO101 real-robot dataset in LeRobot format |

---

## Dataset Statistics

### LIBERO Simulator Benchmark

| Attribute | Value |
|---|---|
| Tasks | 10 |
| Demonstrations | 500 (50 per task) |
| Recorded frames per camera | 62,250 |
| Total recorded frames (both cameras) | 124,500 |
| Frames per demo | 75–197 (mean 124.5) |
| Cameras | `agentview`, `eye_in_hand` |
| RGB resolution | 128 × 128 |
| Mean triplets / frame (agentview) | 42.0 |
| Mean triplets / frame (eye_in_hand) | 16.73 |

### SO101 Real-Robot Dataset

| Attribute | Value |
|---|---|
| Tasks | 5 |
| Demonstrations | 257 (47–53 per task) |
| Total recorded frames (both cameras) | 240,598 |
| VLM eval frames (both cameras) | 8,252 |
| Cameras | `agent_view`, `wrist` |
| Frame rate | 30 FPS (1 frame/sec sampled) |
| Format | LeRobot |

---

## Spatial Ontology

**Objects (LIBERO):** `akita_black_bowl_1`, `akita_black_bowl_2`, `cookies_1`, `glazed_rim_porcelain_ramekin_1`, `plate_1`, `wooden_cabinet_1`, `flat_stove_1`

**Objects (SO101):** `black_bowl`, `red_drawer`, `black_stove`, `cookie`, `white_plate`

**Relations (both):**

| Relation | Description |
|---|---|
| `is_left_of` / `is_right_of` | Lateral world-frame ordering |
| `is_in_front_of` / `is_behind` | Depth world-frame ordering |
| `is_on_top_of` / `is_below_of` | Vertical support / stacking |
| `is_inside` / `contains` | Containment |

---