Whole-body motion planning in dynamic environments
<p><strong>The real world is dynamic.</strong></p> <p>Over the previous decades, we have seen phenomenal advances across the fields of robotics. From the algorithms used to calculate how a robot moves to the hardware on which it is executed, we find ourselves at a poin...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2021
|
| Subjects: |
| _version_ | 1797107606187147264 |
|---|---|
| author | Finean, MN |
| author2 | Havoutis, I |
| author_facet | Havoutis, I Finean, MN |
| author_sort | Finean, MN |
| collection | OXFORD |
| description | <p><strong>The real world is dynamic.</strong></p>
<p>Over the previous decades, we have seen phenomenal advances across the fields of robotics. From the algorithms used to calculate how a robot moves to the hardware on which it is executed, we find ourselves at a point in time where robots appear as though they could realistically be deployed in our homes to carry out complex tasks. And still, we are not there yet.</p>
<p>Despite the tremendous progress that has been made towards enabling such a technological feat, there remains a disconnect between our accomplishments within the robotics community and the applications in which we hope to see them operate. Few works consider the full integration of pipelines and features required to make a robot operate autonomously in real-world environments. As such, fundamental research problems have remained unidentified and thus unaddressed.</p>
<p>In this thesis, we embark on the ambitious, practical goal of enabling a human support robot to perform whole-body manipulation tasks in real-world indoor environments and in the presence of moving obstacles, such as humans. As we strive to achieve this goal, we uncover new challenges and propose novel solutions.</p>
<p>By focusing on trajectory optimisation-based motion planning to achieve reactive motion re-planning, we introduce and develop the concept of predicted composite distance fields for embedding predicted trajectories of moving obstacles into a distance field representation of the environment. We show that this technique can leverage parallel processing to enable a significant performance boost over state-of-the-art GPU-optimised methods of generating distance fields from scratch.</p>
<p>In keeping with the application-driven theme of this thesis, we develop and present a fully integrated motion planning and perception framework. After comparing the different approaches presented in the literature, we propose a fusion of the GPU-based perception framework, GPU-Voxels, and our adaptation of the state-of-the-art optimisation-based motion planner, GPMP2. We demonstrate that our framework can accomplish challenging tasks in dynamic, real-world environments on physical robots.</p>
<p>In the course of this research, we identify a previously unexplored and key challenge for mobile robots, 'Trajectory-Constrained Active Gaze Control', and present a state-of-the-art greedy-optimisation based solution. We believe that identifying this problem is a significant milestone in the pursuit of autonomous robots that operate in dynamic environments. By optimising the capabilities of sensors at our disposal, we can provide both more effective collision avoidance and reduce the need for a multitude of robot-mounted sensors.</p>
<p>To provide further robustness to our framework in the presence of dynamic obstacles, and humans in particular, we incorporate a predictive element. To this end, we propose the Receding Horizon And Predictive Gaussian Process Motion Planner 2 (RHAP-GPMP2), which uses our method of predicted composite distance fields to provide pre-emptive motion planning. We further integrate this with a proposed method for human trajectory prediction that combines intention recognition with trajectory optimisation-based motion planning. Our resultant framework poses a new state-of-the-art benchmark for reactive whole-body motion planning in dynamic environments.</p>
<p>To aid the analysis of our trajectory prediction method, we develop and release the Oxford Indoor Human Motion (Oxford-IHM) dataset. We additionally open-source the code developed in this DPhil research in the hope that this may pave the way for further research in not just pushing our theoretical achievements but also making real-world systems.</p> |
| first_indexed | 2024-03-07T07:18:22Z |
| format | Thesis |
| id | oxford-uuid:6da1a59a-caea-45df-95b5-075c9bdc06a3 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T07:18:22Z |
| publishDate | 2021 |
| record_format | dspace |
| spelling | oxford-uuid:6da1a59a-caea-45df-95b5-075c9bdc06a32022-09-08T08:43:40ZWhole-body motion planning in dynamic environmentsThesishttp://purl.org/coar/resource_type/c_db06uuid:6da1a59a-caea-45df-95b5-075c9bdc06a3RoboticsEnglishHyrax Deposit2021Finean, MNHavoutis, I<p><strong>The real world is dynamic.</strong></p> <p>Over the previous decades, we have seen phenomenal advances across the fields of robotics. From the algorithms used to calculate how a robot moves to the hardware on which it is executed, we find ourselves at a point in time where robots appear as though they could realistically be deployed in our homes to carry out complex tasks. And still, we are not there yet.</p> <p>Despite the tremendous progress that has been made towards enabling such a technological feat, there remains a disconnect between our accomplishments within the robotics community and the applications in which we hope to see them operate. Few works consider the full integration of pipelines and features required to make a robot operate autonomously in real-world environments. As such, fundamental research problems have remained unidentified and thus unaddressed.</p> <p>In this thesis, we embark on the ambitious, practical goal of enabling a human support robot to perform whole-body manipulation tasks in real-world indoor environments and in the presence of moving obstacles, such as humans. As we strive to achieve this goal, we uncover new challenges and propose novel solutions.</p> <p>By focusing on trajectory optimisation-based motion planning to achieve reactive motion re-planning, we introduce and develop the concept of predicted composite distance fields for embedding predicted trajectories of moving obstacles into a distance field representation of the environment. We show that this technique can leverage parallel processing to enable a significant performance boost over state-of-the-art GPU-optimised methods of generating distance fields from scratch.</p> <p>In keeping with the application-driven theme of this thesis, we develop and present a fully integrated motion planning and perception framework. After comparing the different approaches presented in the literature, we propose a fusion of the GPU-based perception framework, GPU-Voxels, and our adaptation of the state-of-the-art optimisation-based motion planner, GPMP2. We demonstrate that our framework can accomplish challenging tasks in dynamic, real-world environments on physical robots.</p> <p>In the course of this research, we identify a previously unexplored and key challenge for mobile robots, 'Trajectory-Constrained Active Gaze Control', and present a state-of-the-art greedy-optimisation based solution. We believe that identifying this problem is a significant milestone in the pursuit of autonomous robots that operate in dynamic environments. By optimising the capabilities of sensors at our disposal, we can provide both more effective collision avoidance and reduce the need for a multitude of robot-mounted sensors.</p> <p>To provide further robustness to our framework in the presence of dynamic obstacles, and humans in particular, we incorporate a predictive element. To this end, we propose the Receding Horizon And Predictive Gaussian Process Motion Planner 2 (RHAP-GPMP2), which uses our method of predicted composite distance fields to provide pre-emptive motion planning. We further integrate this with a proposed method for human trajectory prediction that combines intention recognition with trajectory optimisation-based motion planning. Our resultant framework poses a new state-of-the-art benchmark for reactive whole-body motion planning in dynamic environments.</p> <p>To aid the analysis of our trajectory prediction method, we develop and release the Oxford Indoor Human Motion (Oxford-IHM) dataset. We additionally open-source the code developed in this DPhil research in the hope that this may pave the way for further research in not just pushing our theoretical achievements but also making real-world systems.</p> |
| spellingShingle | Robotics Finean, MN Whole-body motion planning in dynamic environments |
| title | Whole-body motion planning in dynamic environments |
| title_full | Whole-body motion planning in dynamic environments |
| title_fullStr | Whole-body motion planning in dynamic environments |
| title_full_unstemmed | Whole-body motion planning in dynamic environments |
| title_short | Whole-body motion planning in dynamic environments |
| title_sort | whole body motion planning in dynamic environments |
| topic | Robotics |
| work_keys_str_mv | AT fineanmn wholebodymotionplanningindynamicenvironments |