Research Interests

TL;DR: My research develops hybrid neuro-symbolic methods for achieving general-purpose mobile manipulation in robotic assistants. This approach overcomes the brittle generalization and data inefficiency of pure learning methods (like BC) by integrating their strengths with the robustness of classical symbolic planning, focusing on data-efficient and language-conditioned solutions for long-horizon tasks.

------

The growing demand for personalized assistance for an aging and increasingly busy population necessitates the development of general-purpose robotic assistants capable of executing a wide variety of tasks in both home and workplace environments. Think Rosey from The Jetsons. Achieving this goal requires robots that can reliably understand and execute complex, open-vocabulary mobile manipulation tasks based on natural language instructions. This is the ultimate objective of my research.

My research philosophy is driven by a commitment to pragmatism, utilizing the most effective methodologies available, rather than being strictly confined to a single paradigm.

• Critique of Pure Learning: Recent popularity in data-driven methods, such as Imitation Learning (IL), particularly Behavior Cloning (BC), offers benefits for specific domains like tabletop and contact-rich manipulation. However, these methods exhibit significant limitations for general-purpose assistance, including brittle generalization, extreme data requirements, and limited success in mobile manipulation. Similarly, pure Reinforcement Learning (RL) requires extensive sample collection, often relying on simulators that pose a challenging sim-to-real transfer problem.

• Revisiting Classical Methods: Conversely, classical approaches like symbolic planning offer robustness and guaranteed task completion, provided that accurate environmental and task models are supplied. While historically criticized for a lack of environmental flexibility and the need for expert human modeling, their inherent reliability is a compelling advantage when compared to the generalization failures of current learning-based techniques.

Given the complementary strengths and weaknesses across these approaches, my research focuses on developing hybrid methodologies that strategically combine the best elements of both classical and learning-based paradigms, specifically, neuro-symbolic modular systems, to achieve the required robustness and generalizability.

One of my core interests lies in developing data-efficient methods that generalize effectively to long-horizon mobile manipulation tasks.

My work presented at IROS 2025 demonstrated that state-of-the-art BC mobile manipulation methods fail to generalize effectively to novel environments and previously unseen long-horizon tasks using limited training data. Crucially, a hybrid neuro-symbolic method, developed without requiring any real-world robot data, significantly outperformed the BC baselines.

This finding has motivated a shift in my research focus toward the exploration of modular neuro-symbolic techniques that integrate the strengths of both planning and learning. Current projects involve various strategies for combining these elements. Furthermore, I am deeply interested in improving fundamental components of the "neuro" aspect itself, such as enhancing the data efficiency of Vision-Language-Actions (VLA) models and developing better methods for handling long-term memory and contextual reasoning for language-conditioned robotic systems.