Robotic object rearrangement is a core capability for manipulation in human-centric environments, where robots must interact with diverse objects under clutter, occlusions, and dynamic scene changes.
These settings demand decision making under partial...
Robotic object rearrangement is a core capability for manipulation in human-centric environments, where robots must interact with diverse objects under clutter, occlusions, and dynamic scene changes.
These settings demand decision making under partial observability, combinatorial multi-object interactions, and long-horizon objectives, while respecting geometric, kinematic, and safety constraints.
This dissertation develops a unified framework for vision-based hierarchical policy learning with learning-based action primitives, spanning (i) goal-conditioned tabletop rearrangement with explicit target configurations, (ii) goal-free tabletop tidying under implicit organizational preferences, and (iii) indoor humanoid loco-manipulation that requires coordinating navigation and interaction in large 3D environments.
A central thesis of this work is that directly learning end-to-end whole-body control is often inefficient and brittle for rearrangement, particularly for high-DoF robots and contact-rich tasks.
Instead, we construct learning-based action primitives that encapsulate meaningful interaction skills and learn high-level decision-making policies that select and parameterize these primitives from visual observations.
Building on this perspective, the dissertation presents four instantiations of hierarchical action primitives and the corresponding learning algorithms.
First, we study goal-conditioned rearrangement using non-prehensile pushing primitives.
We introduce a vision-based decomposed Q-learning framework that parameterizes pushes in image space and learns value functions for selecting effective push actions under partial observability.
This formulation improves data efficiency by structuring the action space around a manipulation primitive rather than raw joint control.
Second, we improve robustness and generalization by incorporating geometric scene structure.
We propose signed distance field (SDF)-based scene representations and object-centric scene graphs that encode object geometry and spatial relations.
Based on this representation, we develop an SDF-based push primitive and a Q-learning method that generalizes across diverse object configurations and varying numbers of objects in cluttered tabletop environments.
Third, we address goal-free tabletop tidying, where specifying a single target arrangement is impractical and success is defined by an implicit notion of organization.
We construct the Tabletop Tidying Up (TTU) dataset and learn the Tidiness Score, a visual metric that quantifies the degree of tidiness across diverse clutter patterns and environments.
We then define grasp-parameterized pick-and-place primitives using predicted 6-DoF grasping points, train a tidying policy with offline reinforcement learning, and integrate the learned policy and tidiness score into a Monte Carlo Tree Search (MCTS) planner to enable long-horizon decision making for tidying.
Finally, we extend the action-primitive hierarchy to indoor humanoid loco-manipulation.
We present a hierarchical whole-body control framework that couples a vision-language-action (VLA) policy for upper-body manipulation with a separate locomotion policy for stable lower-body control.
The high-level policy outputs high-level locomotion commands (e.g., linear velocity and target yaw) and manipulation actions, while a mode gate mediates between locomotion and manipulation-dominant behaviors across different phases of a task, enabling scalable coordination for large-space rearrangement.
Overall, this dissertation demonstrates that vision-based hierarchical policy learning with learning-based action primitives-combining structured representations, value-based learning, offline policy learning, and planning-provides a principled and scalable approach to robust object rearrangement across tasks, embodiments, and unstructured environments.