Register now After registration you will be able to apply for this opportunity online.
Perceptive Arm Motion Planning and Control for Heavy Construction Machine Tasks
In this work we would utilize reinforcement learning, neural network actuator modeling, and perception for the control and arm motion planning of a 75ton excavator with a free-swinging joint.
The project will be in collaboration with LIEBHERR, a german company building excavators and other construction machines.
Recent advances in the field of robotics led to a full body of new opportunities regarding their utilization. One example for this is the automation of industrial material handling, e.g. in harbors or sorting facilities.
In this project we would like to develop a solution for autonomous material handling on large construction machines, such as the 75 ton LIEBHERR LH80 (as shown in the image). One field of application of these machines is the loading and unloading of ships. These tasks require precise motions without hitting into obstacles. The hydraulic actuators of these machines as well as the free-swinging end effector joint further complicate the execution of these tasks, rendering it only possible for very experienced operators.
Building on prior work done in our lab, the goal of this thesis is to build up a pipeline for the control and collision-free motion planning of the excavator’s arm and cabin in 3D. Due to the hard-to-model hydraulic actuators, in this work we want to utilize reinforcement learning algorithms for the actuator control [1]. During operation, the swinging of the end-effector should be reduced as much as possible during the execution of the plans, either by shaping the horizontal velocity profiles, or by adding it as additional constraints into an optimization problem. As an alternative, the whole control could be learned in a single reinforcement learning policy. Depending on the scope of the project (master thesis or semester project), the project will also include the processing of perceptive information; 3D sensing in the form of LiDAR scans as well as semantic segmentation in camera images. The latter is particularly important to classify and detect the material to be handled, and to allow for a safe operation in presence of (dynamic) objects.
While sampling- and optimization-based methods are widely used for the motion planning of current robotic applications, they are not always able to handle both kinodynamic considerations and arbitrary-shaped obstacles in real-time. Similar to previous research at RSL [2], we would like to combine the advantages of both classes of algorithms to achieve efficient and feasible motion planning. The efficiency and robustness of these methods in real and complex environments with limited map information needs to be further evaluated.
This project will be conducted in collaboration with LIEBHERR, a german company building construction machines. The first tests will be conducted on a LIEBHERR LH40 machine, with the goal to extend it to LH80 in the future.
Recent advances in the field of robotics led to a full body of new opportunities regarding their utilization. One example for this is the automation of industrial material handling, e.g. in harbors or sorting facilities.
In this project we would like to develop a solution for autonomous material handling on large construction machines, such as the 75 ton LIEBHERR LH80 (as shown in the image). One field of application of these machines is the loading and unloading of ships. These tasks require precise motions without hitting into obstacles. The hydraulic actuators of these machines as well as the free-swinging end effector joint further complicate the execution of these tasks, rendering it only possible for very experienced operators.
Building on prior work done in our lab, the goal of this thesis is to build up a pipeline for the control and collision-free motion planning of the excavator’s arm and cabin in 3D. Due to the hard-to-model hydraulic actuators, in this work we want to utilize reinforcement learning algorithms for the actuator control [1]. During operation, the swinging of the end-effector should be reduced as much as possible during the execution of the plans, either by shaping the horizontal velocity profiles, or by adding it as additional constraints into an optimization problem. As an alternative, the whole control could be learned in a single reinforcement learning policy. Depending on the scope of the project (master thesis or semester project), the project will also include the processing of perceptive information; 3D sensing in the form of LiDAR scans as well as semantic segmentation in camera images. The latter is particularly important to classify and detect the material to be handled, and to allow for a safe operation in presence of (dynamic) objects.
While sampling- and optimization-based methods are widely used for the motion planning of current robotic applications, they are not always able to handle both kinodynamic considerations and arbitrary-shaped obstacles in real-time. Similar to previous research at RSL [2], we would like to combine the advantages of both classes of algorithms to achieve efficient and feasible motion planning. The efficiency and robustness of these methods in real and complex environments with limited map information needs to be further evaluated.
This project will be conducted in collaboration with LIEBHERR, a german company building construction machines. The first tests will be conducted on a LIEBHERR LH40 machine, with the goal to extend it to LH80 in the future.
- Literature review on planning and control algorithms
- Real-world data collection on the real machine
- Modeling of the cabin turn joint with the help of a neural network actuator model
- Development of a RL controller
- Implementing the planning algorithm and benchmark performance of existing methods
- Optional: Including perception for collision detection and environment segmentation
- Optional: real-world evaluation tests and experiments
- Literature review on planning and control algorithms - Real-world data collection on the real machine - Modeling of the cabin turn joint with the help of a neural network actuator model - Development of a RL controller - Implementing the planning algorithm and benchmark performance of existing methods
- Optional: Including perception for collision detection and environment segmentation - Optional: real-world evaluation tests and experiments
- Highly motivated for the topic
- Programming experience (C++/Python)
- Knowledge in planning, optimization, and robot dynamics is a plus
- Highly motivated for the topic - Programming experience (C++/Python) - Knowledge in planning, optimization, and robot dynamics is a plus
- Pascal Egli (pasegli@ethz.ch)
- Fang Nan (fannan@ethz.ch)
- Filippo Spinelli (fspinelli@ethz.ch)
- Pascal Egli (pasegli@ethz.ch) - Fang Nan (fannan@ethz.ch) - Filippo Spinelli (fspinelli@ethz.ch)
Liebherr_student_project.pdf
