THESIS
2021
1 online resource (xvi, 118 pages) : illustrations (some color)
Abstract
Autonomous systems evolve rapidly in daily life, from hard-programmed industrial robotic
arms to intelligent household service robots, from low-level driving assistance to high-level
fully self-driving. To accomplish complex tasks with a higher level of autonomy, the capabilities
of efficiently planning actions and accurately reasoning futures is crucial.
In this thesis, I present motion prediction and planning approaches under interactive scenarios.
The first part refers to planning for robotic manipulation under physical interaction. For
robustly grasping arbitrary objects, we propose to optimize fingertip surface and plan grasps by
geometry matching, which improves grasp stability and robustness by maximizing grasp contact
areas. The grasp planning method is further applied to enable...[
Read more ]
Autonomous systems evolve rapidly in daily life, from hard-programmed industrial robotic
arms to intelligent household service robots, from low-level driving assistance to high-level
fully self-driving. To accomplish complex tasks with a higher level of autonomy, the capabilities
of efficiently planning actions and accurately reasoning futures is crucial.
In this thesis, I present motion prediction and planning approaches under interactive scenarios.
The first part refers to planning for robotic manipulation under physical interaction. For
robustly grasping arbitrary objects, we propose to optimize fingertip surface and plan grasps by
geometry matching, which improves grasp stability and robustness by maximizing grasp contact
areas. The grasp planning method is further applied to enable unmanned aerial vehicles
with the capability of making and stabilizing contacts with the environment, to reduce power
consumption and extend the operation time while staying at heights. For efficiently sorting
multiple objects, we employ Monte Carlo tree search with a learned policy as the planning
framework, which is capable of reliably sorting large numbers of objects under complex multi-contact
and modeling inaccuracies. Furthermore, to facilitate flexible and safe motion planning
for autonomous driving under social interaction, the second part highlights motion prediction in
dynamic driving scenarios. A planning-informed prediction method is introduced to tackle the
multi-agent prediction task. By informing the prediction network with the controllable agent’s
planning process, it achieves state-of-the-art prediction accuracy on highway datasets NGSIM
and HighD. Towards accurate and feasible trajectory prediction under explicit constraints and
imperfect tracking, we propose a novel prediction framework, which guarantees trajectory feasibility by exploiting a model-based planner to produce trajectories under constraints and enables accurate multimodal prediction by employing a learning-based evaluator to select trajectories.
The framework shows superiority in prediction accuracy, feasibility, robustness under imperfect
tracking, and achieves 1st place on the Argoverse Motion Forecasting Leaderboard.
Post a Comment