File size: 2,890 Bytes
44e80ff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
task_categories:
- robotics
- reinforcement-learning
language:
- en
tags:
- autonomous-navigation
- robot-navigation
- visual-navigation
- sim-to-real
- igibson
- turtlebot3
- rgb
- lidar
---

# Geometry-guided Representation for Autonomous Navigation

## Overview

The **Geometry-guided Representation for Autonomous Navigation** (_GRAN_) dataset is a collection of simulated robot trajectories designed to study **scene transfer** in autonomous navigation from visual observations.

The dataset contains trajectories collected by a TurtleBot3 robot in the  [iGibson](https://github.com/StanfordVL/iGibson) simulation environment. 
These trajectories are collected across multiple object configurations and visually different environments (background and floor), enabling the study of robust representation learning for vision-based navigation policies to generalize across changes in the appearence of the environment.

## Dataset Composition & Structure

The dataset is composed of:

- **2 rooms** simulated in the iGibson environment;
- **10** different **object settings** per room;
- **4 agents**, with a full knowledge of the environment, differing by the level of expereince;
- **5 trajectories** collected by each agent.

Each trajectory is a collection of RGB images captured by an onboard camera of the TB3 robot, and instantiated in **9 visually different environments**.

The structure of the dataset:

```bash
GRAN/
└── Room1/                             # Room  
    └── Setting1/                      # Room setting
        ├── 8m/                        #
        ├── 6000000/                   # Agents used for the collection of rollout trajectories.
        ├── 3200000/                   # The level of experience is identified by the number of training steps.
        └── 400000/                    #
            └── episode_0001           # Trajectory
                ├── episode_0001.pkl   # Pandas DataFrame object containing per step additional information (e.g., robot's and target's absolute coordinates, LiDAR readings, etc.)
                └──  augmented_results # Trajectory of images collected in the 9 visually different environments
```

## Intended Use

The dataset is intended for research on:

- Robust Representation Learning
- Representation Learning guided by Privileged Information
- Navigation Policy Learning from visual observations
- Scene Transfer of Navigation Policies

## Citation

If you use this dataset in your research, please cite the associated work:

```bibtex 
@article{zhalehmehrabi2026robust,
  title={Robust Scene Transfer for PointGoal Navigation via Privileged Sensor Guided Contrastive Learning},
  author={Zhalehmehrabi, Amirhossein and Tezze, Tiziano and Castelini, Alberto and Farinelli, Alessandro},
  journal={arXiv preprint arXiv:2606.05506},
  year={2026}
}
```