ntnu-arl/unified_autonomy_stack_datasets
Unified Autonomy Stack Datasets This repository contains data released in relation with the Unified Autonomy Stack. Details can be found here Paper: https://arxiv.org/abs/2605.12735 Code: https://ntnu-arl.github.io/unified_autonomy_stack/ The data was collected using the following platforms in both manually piloted and autonomously operated modes: AR-1 (Hornbill): A variant of the RMF-Owl collision-tolerant aerial robot. AR-2 (Magpie): A collision-tolerant aerial robot… See the full description on the dataset page: https://huggingface.co/datasets/ntnu-arl/unified_autonomy_stack_datasets.
Unified Autonomy Stack Datasets
This repository contains data released in relation with the Unified Autonomy Stack. Details can be found here
- Paper: <https://arxiv.org/abs/2605.12735>
- Code: <https://ntnu-arl.github.io/unifiedautonomystack/>
The data was collected using the following platforms in both manually piloted and autonomously operated modes:
- AR-1 (Hornbill): A variant of the RMF-Owl collision-tolerant aerial robot.
- AR-2 (Magpie): A collision-tolerant aerial robot designed to carry the UniPilot module.
- GR-1 (Anymal): An ANYmal D quadruped robot from ANYbotics carrying the UniPilot module.
- UniPilot: A compact hardware-software autonomy payload that can be integrated across diverse robot embodiments to enable autonomous operation in GPS-denied environments. In addition to being carried by AR-2 and GR-1, a handheld (helmet-mounted) variant was used for the
campus_fogsequence.
The following table summarizes the platforms used for each dataset:
Sensor Setup
Data Description
Each sequence is released as a ROS 1 bag named sensors_only.bag containing the raw sensor data collected during the run. The ground truth is provided as a .tum file for the sequences where it is available.
Due to resource constraints on the platforms, the LiDAR is recorded as raw packets instead of the deserialized point clouds. To obtain bags with the point clouds, please view the lidar_packets_to_pointclouds.md.
Topics
The radar, IMU, and static transforms are common to all platforms; only the LiDAR and camera topics differ. Bags may additionally contain: flight-controller telemetry under /mavros/* (including GNSS as sensor_msgs/NavSatFix and mavros_msgs/GPSRAW where a fix was available), hardware time-synchronization triggers and status under /sensor_sync_node/*, further VN-100 streams (temperature, filtered IMU), and per-sensor host-receive timestamps under /<sensor>/ros_time_now.
Common to all platforms
LiDAR
Recorded as raw packets corresponding to a 10 Hz point cloud (one LiDAR per platform):
Cameras
sensor_msgs/CompressedImage at 20 Hz:
On fyllingsdalen_tunnel, the AR-1 camera and radar were recorded at 25 Hz (the radar chirp configuration was changed for high-speed flight).Calibration
Intrinsics
Camera intrinsics are provided in the calibration folder, with the same intrinsics for all sequences recorded on a given platform.
Extrinsics
The transforms between sensors are provided below. The transform $T{AB}$ transforms a point from frame $B$ to frame $A$ as $pA = T{AB} * pB$, where the point is represented in homogeneous coordinates. All transforms are provided in the format: x, y, z, qx, qy, qz, qw.
AR-1 (Hornbill):
T_imu_lidar: [0.0166, 0.02158, 0.03375, 0, 0, 0, 1]
T_imu_radar: [0.077, 0.016, -0.063, 0.963, -0.021, -0.265, 0.021]
T_imu_cam: [0.0725765278611583, 0.018936068067624674, -0.03560091123164558, 0.5543417213240229, -0.5433799916950063, 0.44199772312966146, 0.4495347076402992]
AR-2 / GR-1 (UniPilot):
T_imu_lidar: [-0.06605, -0.01878, 0.034, 0.707, 0.00, -0.707, 0.00]
T_imu_radar: [0.07717380907196035, -0.0479741168664902, 0.006770362043579366, 0.01761050968654745, 0.25032180256367187, -0.017699746191127304, 0.9679631109162291]
UniPilot handheld:
T_imu_lidar: [-0.11484, -0.01878, 0.035, 0.3799282, -0.5963678, -0.3799282, 0.5963678]
T_imu_radar: [0.08373, 0.0213, 0.02691, 0.6830127, 0.1830128, -0.1830128, 0.6830127]
T_imu_cam_front: [0.083743715, 0.000235985, 0.005836153, -0.608415182, 0.616574414, -0.353450887, 0.353184694]
T_imu_cam_left: [-0.021764534, 0.045545798, 0.023474280, -0.706455147, 0.000556152, -0.002366216, 0.707753642]
T_imu_cam_right: [-0.025516567, -0.089169163, 0.024118153, 0.008477527, 0.715010265, -0.699047196, -0.004633580]
Ground Truth
Where available, ground truth is provided as a TUM-format trajectory file (timestamp tx ty tz qx qy qz qw) alongside the bag.
- Tunnels (`fyllingsdalen_tunnel`, `runehamar_tunnel/hornbill`): generated by fusing the tracking of a Leica GRZ101 mini-prism (mounted on AR-1) by a Leica MS60 MultiStation with the onboard IMU, in an offline Levenberg-Marquardt optimization.
- `campus_fog` and `frozen_lake`: GNSS was available, so ground truth is created using a GNSS-augmented visual bundle adjustment optimization with Pix4DMatic.
Citation
If you use this data in your research, please cite the following publication:
@misc{dharmadhikari2026unifiedautonomystackblueprint,
title={The Unified Autonomy Stack: Toward a Blueprint for Generalizable Robot Autonomy},
author={Mihir Dharmadhikari and Nikhil Khedekar and Mihir Kulkarni and Morten Nissov and Martin Jacquet and Angelos Zacharia and Marvin Harms and Albert Gassol Puigjaner and Philipp Weiss and Kostas Alexis},
year={2026},
eprint={2605.12735},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2605.12735},
}