EnvoDat: A Large-Scale Multisensory Dataset for Robotic Spatial
Awareness and Semantic Reasoning in Heterogeneous Environments.
Linus Nwankwo, Bjoern Ellensohn, Vedant Dave, Peter Hofer, Jan Forstner, Marlene Villneuve, Robert Galler and Elmar Rueckert
EnvoDat is a large-scale, multisensory dataset for advancing robot spatial awareness and semantic reasoning in heterogeneous environments. EnvoDat is captured in diverse environments, under several environmental conditions, including dynamic entities, high illumination, shadow formations, fog, rain, and zero and partial visibility conditions at different times of the day. EnvoDat is introduced to support the development of robust and ecocentric SLAM and supervised learning algorithms, to push the boundaries of robot autonomy, spatial awareness and scene understanding across multi-structural, unstructured natural, in-outdoor, and subterranean environments.
Poor visibility & feature-sparse sub tunnel
Open street, feature sparse & dynamic
Textureless, sparse & artificial lightening
Dense canopies & elevated terrain
Partial visibility with fluctuating lightening
Soft, feature-sparse & varying lightening
Soft, feature-sparse and sloppy terrain
Featureless, textureless & artificial lightening
Granular & elevated terrain
Wet floor, high symmetric & dynamic
Degraded floor & day lightening
Partial vis., featureless & artificial lightening
Multiple occlusions
Rain, fog & dynamic entities
Rocky & uneven terrain
High symmetric tunnel
Dense canopies & sloppy terrain
Sandy, feature-free & narrow tunnel
Varying illumination & dynamic entities
Vegetated & shadow formation
Vegetated scene
Vegetated path
High illumination
Recorded in 13 challenging indoor, outdoor and underground environments.
Captured under several environmental conditions e.g., dynamic entities, high illumination, shadow formations, foggy, rain, zero and partial visibility conditions.
Intermittently collected between November 2022 and September 2024 at different times.
ROS AproximateTime Policy synchronized data from 10 sensing modalities.
1.9+ TB of data provided in different formats suitable for:
benchmarking SLAM, perception, and supervised learning algorithms.
multimodal reasoning, grounding, and downstream applications e.g., fine-tuning open-source multimodal vision models.
Richly fine-grained layered polygon-based annotations containing:
89K annotations at an average of 10.3 annotations per frame.
82 object and terrain classes.
ICRA2025VideoSummary.mp4If you use this work in your research, please cite it using the following BibTeX entry:
@INPROCEEDINGS{11127594,
author={Nwankwo, Linus and Ellensohn, Björn and Dave, Vedant and Hofer, Peter and Forstner, Jan and Villneuve, Marlene and Galler, Robert and Rueckert, Elmar},
booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)},
title={EnvoDat: A Large-Scale Multisensory Dataset for Robotic Spatial Awareness and Semantic Reasoning in Heterogeneous Environments},
year={2025},
volume={},
number={},
pages={153-160},
keywords={Simultaneous localization and mapping;Annotations;Heuristic algorithms;Supervised learning;Semantics;Benchmark testing;Cognition;Spatial databases;Robustness;Sensors},
doi={10.1109/ICRA55743.2025.11127594}}
This project has received funding from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - No #430054590 (TRAIN).
