SubT-MRS

depth · video · SubT-MRS is an extremely challenging real-world dataset designed to push SLAM toward all-weather, perceptually-degraded environments. It comprises roughly five years of data: three years from the DARPA Subterranean (SubT) Challenge (2019-2021) plus two additional years of diverse environments (2022-2023). The collection covers over 2000 hours and 300+ kilometers of terrain across more than 30 diverse scenes, including subterranean caves, tunnels, urban areas, long structureless corridors, mixed indoor/outdoor settings, off-road terrain, deserts, forests and bushlands. It captures extreme conditions such as dense fog, dust, smoke, heavy snow, darkness and varying illumination. Data is collected on multiple robot platforms: unmanned ground vehicles (UGV1/2/3), RC cars (RC1/2/7), legged robots (Boston Dynamics Spot), aerial robots (UAV/drone) and handheld devices, in a multi-robot configuration. Each platform carries hardware time-synchronized multimodal sensors: up to 4 RGB cameras (plus fisheye), one LiDAR, one IMU and one thermal camera. The dataset is distributed in ROS bag format and an extracted folder format, with ground-truth trajectories provided in TUM format (timestamp x y z q_x q_y q_z q_w) and initialization poses. It is organized into subsets including the SubT-MRS main track, a Sensor Fusion extension, a TartanAir LiDAR track (which adds depth and semantic segmentation) and the SuperLoc subset, totaling roughly 25 sequences across the released tracks. The associated CVPR 2024 paper ('SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments', Zhao et al.) also introduces accuracy and robustness evaluation tracks with novel robustness metrics. Released under CC BY 4.0 and hosted on the Super Odometry platform by CMU's AirLab (Robotics Institute).

SubT-MRS is an extremely challenging real-world dataset designed to push SLAM toward all-weather, perceptually-degraded environments. It comprises roughly five years of data: three years from the DARPA Subterranean (SubT) Challenge (2019-2021) plus two additional years of diverse environments (2022-2023). The collection covers over 2000 hours and 300+ kilometers of terrain across more than 30 diverse scenes, including subterranean caves, tunnels, urban areas, long structureless corridors, mixed indoor/outdoor settings, off-road terrain, deserts, forests and bushlands. It captures extreme conditions such as dense fog, dust, smoke, heavy snow, darkness and varying illumination. Data is collected on multiple robot platforms: unmanned ground vehicles (UGV1/2/3), RC cars (RC1/2/7), legged robots (Boston Dynamics Spot), aerial robots (UAV/drone) and handheld devices, in a multi-robot configuration. Each platform carries hardware time-synchronized multimodal sensors: up to 4 RGB cameras (plus fisheye), one LiDAR, one IMU and one thermal camera. The dataset is distributed in ROS bag format and an extracted folder format, with ground-truth trajectories provided in TUM format (timestamp x y z q_x q_y q_z q_w) and initialization poses. It is organized into subsets including the SubT-MRS main track, a Sensor Fusion extension, a TartanAir LiDAR track (which adds depth and semantic segmentation) and the SuperLoc subset, totaling roughly 25 sequences across the released tracks. The associated CVPR 2024 paper ('SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments', Zhao et al.) also introduces accuracy and robustness evaluation tracks with novel robustness metrics. Released under CC BY 4.0 and hosted on the Super Odometry platform by CMU's AirLab (Robotics Institute).

SubT-MRS is an extremely challenging real-world dataset designed to push SLAM toward all-weather, perceptually-degraded environments. It comprises roughly five years of data: three years from the DARPA Subterranean (SubT) Challenge (2019-2021) plus two additional years of diverse environments (2022-2023). The collection covers over 2000 hours and 300+ kilometers of terrain across more than 30 diverse scenes, including subterranean caves, tunnels, urban areas, long structureless corridors, mixed indoor/outdoor settings, off-road terrain, deserts, forests and bushlands. It captures extreme conditions such as dense fog, dust, smoke, heavy snow, darkness and varying illumination. Data is collected on multiple robot platforms: unmanned ground vehicles (UGV1/2/3), RC cars (RC1/2/7), legged robots (Boston Dynamics Spot), aerial robots (UAV/drone) and handheld devices, in a multi-robot configuration. Each platform carries hardware time-synchronized multimodal sensors: up to 4 RGB cameras (plus fisheye), one LiDAR, one IMU and one thermal camera. The dataset is distributed in ROS bag format and an extracted folder format, with ground-truth trajectories provided in TUM format (timestamp x y z q_x q_y q_z q_w) and initialization poses. It is organized into subsets including the SubT-MRS main track, a Sensor Fusion extension, a TartanAir LiDAR track (which adds depth and semantic segmentation) and the SuperLoc subset, totaling roughly 25 sequences across the released tracks. The associated CVPR 2024 paper ('SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments', Zhao et al.) also introduces accuracy and robustness evaluation tracks with novel robustness metrics. Released under CC BY 4.0 and hosted on the Super Odometry platform by CMU's AirLab (Robotics Institute).

SubT-MRS is an extremely challenging real-world dataset designed to push SLAM toward all-weather, perceptually-degraded environments. It comprises roughly five years of data: three years from the DARPA Subterranean (SubT) Challenge (2019-2021) plus two additional years of diverse environments (2022-2023). The collection covers over 2000 hours and 300+ kilometers of terrain across more than 30 diverse scenes, including subterranean caves, tunnels, urban areas, long structureless corridors, mixed indoor/outdoor settings, off-road terrain, deserts, forests and bushlands. It captures extreme conditions such as dense fog, dust, smoke, heavy snow, darkness and varying illumination. Data is collected on multiple robot platforms: unmanned ground vehicles (UGV1/2/3), RC cars (RC1/2/7), legged robots (Boston Dynamics Spot), aerial robots (UAV/drone) and handheld devices, in a multi-robot configuration. Each platform carries hardware time-synchronized multimodal sensors: up to 4 RGB cameras (plus fisheye), one LiDAR, one IMU and one thermal camera. The dataset is distributed in ROS bag format and an extracted folder format, with ground-truth trajectories provided in TUM format (timestamp x y z q_x q_y q_z q_w) and initialization poses. It is organized into subsets including the SubT-MRS main track, a Sensor Fusion extension, a TartanAir LiDAR track (which adds depth and semantic segmentation) and the SuperLoc subset, totaling roughly 25 sequences across the released tracks. The associated CVPR 2024 paper ('SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments', Zhao et al.) also introduces accuracy and robustness evaluation tracks with novel robustness metrics. Released under CC BY 4.0 and hosted on the Super Odometry platform by CMU's AirLab (Robotics Institute).

depth · SubT-MRS is an extremely challenging real-world dataset designed to push SLAM toward all-weather, perceptually-degraded environments. It comprises roughly five years of data: three years from the DARPA Subterranean (SubT) Challenge (2019-2021) plus two additional years of diverse environments (2022-2023). The collection covers over 2000 hours and 300+ kilometers of terrain across more than 30 diverse scenes, including subterranean caves, tunnels, urban areas, long structureless corridors, mixed indoor/outdoor settings, off-road terrain, deserts, forests and bushlands. It captures extreme conditions such as dense fog, dust, smoke, heavy snow, darkness and varying illumination. Data is collected on multiple robot platforms: unmanned ground vehicles (UGV1/2/3), RC cars (RC1/2/7), legged robots (Boston Dynamics Spot), aerial robots (UAV/drone) and handheld devices, in a multi-robot configuration. Each platform carries hardware time-synchronized multimodal sensors: up to 4 RGB cameras (plus fisheye), one LiDAR, one IMU and one thermal camera. The dataset is distributed in ROS bag format and an extracted folder format, with ground-truth trajectories provided in TUM format (timestamp x y z q_x q_y q_z q_w) and initialization poses. It is organized into subsets including the SubT-MRS main track, a Sensor Fusion extension, a TartanAir LiDAR track (which adds depth and semantic segmentation) and the SuperLoc subset, totaling roughly 25 sequences across the released tracks. The associated CVPR 2024 paper ('SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments', Zhao et al.) also introduces accuracy and robustness evaluation tracks with novel robustness metrics. Released under CC BY 4.0 and hosted on the Super Odometry platform by CMU's AirLab (Robotics Institute).

Open · CC-BY-4.0 · custom · rosbag