Learning target-driven robotic manipulation in constrained clutter
Robotic manipulation enables robots to interact with and transform their physical environment. Although modern robots perform well in structured industrial settings, robust manipulation in cluttered and constrained environments remains a significant challenge. In such scenarios, robots must reason about complex object interactions, contact dynamics, and limited workspace while executing precise motions.
In his thesis, Yongliang Wang investigates learning-based approaches for enabling target-driven robotic manipulation in constrained and cluttered environments. Wang focuses on improving manipulation efficiency, robustness, and generalization through reinforcement learning, deep reinforcement learning, and diffusion-based policy learning.
First, Wang develops a reinforcement learning framework to improve trajectory planning for robotic manipulators. By integrating vision-based task-space planning with a reinforcement learning controller in joint space, the system enables efficient obstacle avoidance and reactive motion generation for manipulators operating in complex environments. Second, Wang proposes several deep reinforcement learning frameworks to address challenging manipulation tasks where direct grasping is difficult or impossible. These include learning push–grasp coordination for extracting objects from dense clutter and developing dual-arm coordination strategies for manipulating large objects beyond the capability of a single arm. Finally, Wang explores diffusion-based policies for generalizable manipulation. Using scooping as a representative contact-rich task, he introduces a simulation-driven learning framework and dataset to train object-centric policies that can generalize across diverse objects and environments.
Together, these contributions advance the development of intelligent robotic manipulation systems capable of operating reliably in cluttered and constrained real-world environments.