The field has long pursued two parallel yet complementary approaches to robotic manipulation. Classical planning methods explicitly model system dynamics and interactions, enabling scalability, adaptability, and strong formal guarantees—but confining applicability to well-understood systems. In contrast, recent work abandons explicit modeling in favor of data-driven flexibility, capturing interaction dynamics that are difficult to model but struggling to generalize when data is scarce or task distributions shift. Even though the strengths of one mirror the weaknesses of the other, only few efforts have sought principled ways to combine them within a unified framework for multi-robot manipulation.

This thesis targets the boundary between explicit planning and learned policy synthesis, developing algorithms that plan what can be modeled and learn what cannot. By exploiting structure and identifying where learning is tractable, we develop methods for multi-robot manipulation that generalize across team sizes and robot embodiments while retaining the flexibility to learn hard-to-model components.

We begin at the structured end of the spectrum, addressing scalable motion planning for multi-robot-arm systems. After introducing two algorithms for labeled settings–where each robot has an assigned goal–we turn to our first proposed work: an anonymous multi-arm motion planner that generates coordinated trajectories under interchangeable-goal formulations. This capability is essential when policies specify contact objectives without prescribing which robot should execute them.

From this structured foundation, we dive into the boundary between planning and learning.
We develop two methods that compose simple, data-efficient policies through lightweight planning—one for coordination under implicit objectives and another for planar collaborative manipulation—and propose our second research body direction: extending these ideas to realistic 3D domains involving contact-rich manipulation of large rigid objects. This work determines when analytical models suffice, when learning is necessary, and seeks to invoke learned models where they perform well (i.e., are in-distribution), defining a systematic interface between explicit planning and learned interaction dynamics.

Together, these contributions reveal how structure and learning can coexist: flexible and generalizable multi-robot manipulation emerges not from choosing between planning and learning, but from combining them.

• Prof. Jiaoyang Li (co-chair)
• Prof. Maxim Likhachev (co-chair)
• Prof. Andrea Bajcsy
• Prof. Yilun Du (Harvard University)