Hybrid Motion Planning and Control for Legged-Wheeled Robots: Application to Walking Excavators
Open access
Author
Date
2023Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
While slowly finding their way into human-engineered environments, deploying robots in natural environments remains challenging today. Automation is especially lacking for large-scale hydraulic machinery, which would be indispensable for automating dangerous tasks such as natural disaster responses.
For robotic autonomy, motion planning plays an important role, especially in the presence of obstacles. Overcoming obstacles requires adapting locomotion strategy to the surrounding terrain, a pattern that can be observed in humans and animals. Humans walk on two limbs but can use all four if the situation requires so. The terrain around the robot imposes constraints on limb placement, stability, and contact timing, and accounting for all constraints in a single motion planning problem is demanding. Traditionally, the problem is decomposed into smaller subproblems using simplified models and heuristics, which often cannot capture the coupled dynamics in the system. Hence they often plan motions not fully utilizing the robot's capabilities. Treating the robot as a whole becomes especially important for complex systems such as legged-wheeled robots.
This dissertation extends the locomotion capabilities of legged robots, emphasizing legged excavators. It develops motion planning algorithms utilizing all degrees of freedom for overcoming challenging terrain. We formulate the motion planning problem in a general way for multiple robot types and explore concepts for solving it. To this end, optimization and randomized sampling play a central role in computing global, whole-body motions presented in this thesis.
We work towards increasing legged robotic mobility in a series of five publications. We start by deploying a hydraulic 12-tonne-legged machine in a natural forest for precision harvesting. The mapping, localization, planning, and control systems have been proposed and integrated into an autonomous harvesting solution. The whole-body planner developed later in the thesis is inspired by the shortcomings of the planning system deployed in the forest.
A local optimization-based planner is used to cope with many degrees of freedom on a legged excavator. The proposed terrain-aware planner can compute efficient driving and stepping motion and utilize the arm for locomotion. To allow more flexibility, we only command goal poses and do not stipulate what to do in between, thus giving the optimizer complete freedom.
To cope with the non-convexity of the planning problem, we employ randomized sampling. We shift some of the computation offline in the form of pre-computed roadmaps, which help keep the planning times low. The initial whole-body plan is found by randomized sampling; however, it may still violate some physical system constraints (e.g., wheel rolling constraints). The initial motion plan is then fed to the optimization for refinement. The optimization ensures constraint satisfaction, while the initial plan keeps the optimizer away from bad local minima. Together they compute smooth, global, whole-body plans. We have tested this method on a legged excavator, a legged-wheeled robot, and a legged robot with point feet.
Lastly, a control system for legged excavators is developed. Whole-body motion plan execution is a well-studied topic for more miniature robots such as quadrupeds. However, hardware deployment on full-size hydraulic machinery is lacking. We run the optimization in a receding horizon fashion, which helps to combat drift and tracking errors. The terrain adaptive control system allows the planner to plan on simplified geometries and then adapt to unseen landscapes at execution time. Planning on simplified geometries is essential for deployment since a very accurate map cannot be obtained with current sensors. Small details like steps and roughness produce non-smooth gradients that hamper the optimization convergence. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000615353Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZurichSubject
Motion Planning; Robotics; Control; Autonomous Excavator; Walking excavator; Optimization and Optimal Control; Randomized sampling; RSLOrganisational unit
09570 - Hutter, Marco / Hutter, Marco
02284 - NFS Digitale Fabrikation / NCCR Digital Fabrication
More
Show all metadata
ETH Bibliography
yes
Altmetrics