pipeline_tag: robotics
tags:
- model_hub_mixin
- pytorch_model_hub_mixin
PointGoal Navigation Policy (FrodoBots8K)
PointGoal navigation policy trained with the FrodoBots8K dataset, as presented in the paper Data Scaling for Navigation in Unknown Environments.
Authors
Lauri Suomela, Naoki Takahata, Sasanka Kuruppu Arachchige, Harry Edelman, Joni-Kristian Kämäräinen
Model Description
This model is an end-to-end, map-free visual navigation policy designed for sidewalk robots. It was trained using imitation learning on crowd-sourced data to achieve zero-shot navigation in unknown environments. The research demonstrates that data diversity (geographical variety) is significantly more important than data quantity for real-world generalization.
- Task: Zero-shot PointGoal navigation.
- Generalization: Evaluated across 125 km of autonomous driving in four different countries.
- Key Finding: Doubling geographical locations in training decreases navigation errors by ~15%, while adding more data from existing locations leads to rapid performance saturation.
Details
Architecture
- Type: MLP-BC
- Visual Encoder: Theia encoder
- Action Head: MLP action head
Training Dataset
- Dataset: FrodoBots8K (subset)
- Scale: 1024 total trajectories (~32 hours of data) from 32 distinct geographical locations.
Usage
The model is built for deployment on EarthRover Zero robots using ROS2. For detailed instructions on environment setup, training, and deployment, please refer to the official GitHub repository.
Citation
If you use this model or code in your research, please cite:
@misc{suomela2026data_scaling,
title={Data Scaling for Navigation in Unknown Environments},
author={Suomela, Lauri and Takahata, Naoki and Kuruppu Arachchige, Sasanka and Edelman, Harry and Kämäräinen, Joni-Kristian},
journal={arXiv:2601.09444},
year={2026},
}