Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Model Adaptation for Predictive Control in Aerial Manipulation
In this thesis we aim to improve a control framework to enable aerial physical manipulation of objects with onboard perception. To this end, we combine a sampling-based MPC approach with onboard sensing (e.g., RGBD camera) on a drone in order to update the internal controller model according to the perceived information.
Aerial Manipulators (AM), a new class of Micro Aerial Vehicles (MAVs), are designed to physically interact with their environment. This can include touching surfaces (e.g., for aerial inspection), or actively manipulating objects.
In preliminary works we have developed a control framework that enables an aerial manipulator robot to actively interact with objects and we have shown this in examples like opening a door or turning a valve. The controller is based on a Model Predictive Path Integral (MPPI) approach which computes commands by optimizing over multiple rollouts obtained by a physics simulator (Raisim).
While this approach works well, it is heavily dependent on an accurate description of the environment and its dynamics.
In this thesis, we aim to build upon the existing framework and increase its abilities by adapting the internal model according to perceived information of the environment. This would allow in the future to bring the AM into less controlled environments (e.g., outside of laboratory conditions, using onboard sensing only).
This can be achieved by employing an onboard RGBD camera to analyze the environment and update the internal model in the MPPI controller accordingly. However, also other sensors could be used (e.g., force/torque sensors).
Aerial Manipulators (AM), a new class of Micro Aerial Vehicles (MAVs), are designed to physically interact with their environment. This can include touching surfaces (e.g., for aerial inspection), or actively manipulating objects. In preliminary works we have developed a control framework that enables an aerial manipulator robot to actively interact with objects and we have shown this in examples like opening a door or turning a valve. The controller is based on a Model Predictive Path Integral (MPPI) approach which computes commands by optimizing over multiple rollouts obtained by a physics simulator (Raisim). While this approach works well, it is heavily dependent on an accurate description of the environment and its dynamics.
In this thesis, we aim to build upon the existing framework and increase its abilities by adapting the internal model according to perceived information of the environment. This would allow in the future to bring the AM into less controlled environments (e.g., outside of laboratory conditions, using onboard sensing only). This can be achieved by employing an onboard RGBD camera to analyze the environment and update the internal model in the MPPI controller accordingly. However, also other sensors could be used (e.g., force/torque sensors).
Literature r- eview on MPPI, MPC, robot perception, 3D pointcloud fitting
- Familiarization with existing framework, platform, and senors
- Implementation in simulation
- Implementation and experimental validation on a real platform
- Evaluation of results and comparison with existing methods
Literature r- eview on MPPI, MPC, robot perception, 3D pointcloud fitting - Familiarization with existing framework, platform, and senors - Implementation in simulation - Implementation and experimental validation on a real platform - Evaluation of results and comparison with existing methods
- Knowledge of MPC or RL-based controllers
- Good experience with C++
- Experience with ROS preferable (but not required)
- Experience with performing experiments on real robots desirable
- Knowledge of MPC or RL-based controllers - Good experience with C++ - Experience with ROS preferable (but not required) - Experience with performing experiments on real robots desirable