Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Efficient Submaps and Submap Alignment for Online Path Planning
The goal of this project is to create a fast and accurate 3D submap representation of a local environment from sensor data from a mobile robot with drifting odometry.
Keywords: Autonomous Robots, UAV, Mapping, 3D Reconstruction, Signed Distance Fields.
Under drifting odometry it is crucial to properly align incoming sensor data (e.g. from a stereo camera) to the existing internal map in order to maintain a consistent global map. There are numerous approaches to localize and align local submaps to each other or to a global map, but the most efficient data representation and algorithms are unclear.
Singed Distance Fields (SDFs) are often used for local path planning on-board Unmanned Aerial Vehicles (UAVs). An SDF maps every location in 3D space to a field value which represents the distance to the closest obstacle/surface in the scene. This information is invaluable for path-planning and path-optimization on an autonomous mobile robot.
Typically, SDFs discretize space into a uniform grid of voxels. Thus, every point in space can be represented with the same level of accuracy defined by the grid size. But local sumbaps might benefit from a non-uniform grid with a smaller voxel size near the sensor position and larger voxels further away. This makes sense, if the sensor noise is expected to grow with distance, e.g. depth from a stereo camera.
A non-uniform voxel grid could potentially lead to better submap alignment, due to the higher voxel resolution at close distances, while being computationally less intense due to needing fewer voxels overall.
The goal of this project is to examine different map representations and to implement non-uniform local SDFs, alignment algoithms for these non-uniform SDFs, and to extensively test the chosen approach.
Under drifting odometry it is crucial to properly align incoming sensor data (e.g. from a stereo camera) to the existing internal map in order to maintain a consistent global map. There are numerous approaches to localize and align local submaps to each other or to a global map, but the most efficient data representation and algorithms are unclear.
Singed Distance Fields (SDFs) are often used for local path planning on-board Unmanned Aerial Vehicles (UAVs). An SDF maps every location in 3D space to a field value which represents the distance to the closest obstacle/surface in the scene. This information is invaluable for path-planning and path-optimization on an autonomous mobile robot.
Typically, SDFs discretize space into a uniform grid of voxels. Thus, every point in space can be represented with the same level of accuracy defined by the grid size. But local sumbaps might benefit from a non-uniform grid with a smaller voxel size near the sensor position and larger voxels further away. This makes sense, if the sensor noise is expected to grow with distance, e.g. depth from a stereo camera. A non-uniform voxel grid could potentially lead to better submap alignment, due to the higher voxel resolution at close distances, while being computationally less intense due to needing fewer voxels overall.
The goal of this project is to examine different map representations and to implement non-uniform local SDFs, alignment algoithms for these non-uniform SDFs, and to extensively test the chosen approach.
- **WP1**: Familiarization with state-of-the-art mapping approaches for path-planning and for 3D reconstruction.
- **WP2**: Design an efficient local submap e.g. in the form of a non-uniform SDF.
- **WP3**: Implement the framework and create benchmarks for comparisons with state-of-the-art approaches
- **WP4**: Design and conduct experiments with a UAV to evaluate the selected approach.
- **WP1**: Familiarization with state-of-the-art mapping approaches for path-planning and for 3D reconstruction. - **WP2**: Design an efficient local submap e.g. in the form of a non-uniform SDF. - **WP3**: Implement the framework and create benchmarks for comparisons with state-of-the-art approaches - **WP4**: Design and conduct experiments with a UAV to evaluate the selected approach.
- Interest in Computer Sciences, Robotics and Autonomous Navigation;
- C++ programming experience;
- Experience in mobile robotics, Linux, ROS is beneficial.
- Interest in Computer Sciences, Robotics and Autonomous Navigation; - C++ programming experience; - Experience in mobile robotics, Linux, ROS is beneficial.
Interested students please send CV, Bachelor and Master transcripts to Yves Kompis (ykompis@ethz.ch) and Luca Bartolomei (lbartolomei@ethz.ch). **Do not** apply on Sirop.
Interested students please send CV, Bachelor and Master transcripts to Yves Kompis (ykompis@ethz.ch) and Luca Bartolomei (lbartolomei@ethz.ch). **Do not** apply on Sirop.