Background
Care environments increasingly require support for repetitive, time-sensitive, and safety-critical tasks, including transporting supplies, retrieving medications, and handling daily necessities. Robots could reduce workload for caregivers and improve service consistency, but practical deployment remains difficult because care environments are dynamic, objects are diverse, and real-world robot training is costly.
This project is motivated by three needs. First, care robots must generalize to unseen objects and tasks without frequent retraining. Second, they must transfer knowledge from simulation or other domains to real hospitals and homes, where collecting data is expensive and risky. Third, they must act in ways that are predictable and human-like so that caregivers and patients can collaborate with them comfortably and safely.
Research Objectives
The main goal is to develop learning-based robotic systems that can assist care work in homes, hospitals, or care centers.
The first goal is to achieve zero-shot manipulation, allowing robot arms to handle unfamiliar care-related objects without additional task-specific training.
The second goal is to improve transfer from simulation or other source domains to real robotic platforms, reducing the cost and risk of real-world data collection.
The third goal is to make robot behavior more human-like, predictable, and acceptable to caregivers and patients. Together, these goals aim to support practical care applications such as delivering medications, moving daily necessities, and assisting routine logistics tasks while maintaining adaptability, efficiency, and natural interaction with people.
Methods
The methodology consists of three connected components. First, the project will study zero-shot RL generalization for robotic manipulation. Robot arms will undergo simulation pre-training using reward-free pre-training via forward-backward representations, so that they can acquire general-purpose manipulation capabilities before task-specific rewards are introduced. This will support transfer to real-world tasks involving unseen objects and new care scenarios. Second, the project will develop cross-domain reinforcement learning methods to bridge the gap between source and target domains. This includes learning inter-domain mappings when states, actions, robot morphology, or embodiment differ. The approach will use cross-domain Bellman consistency to guide mapping learning, together with hybrid critics and adaptive weighting to balance source-domain and target-domain value functions. Third, the project will incorporate human-like RL. Human demonstrations will be distilled into macro actions using macro action quantization, and robot decision-making will be formulated as trajectory optimization so that action sequences better align with human behavior. These components will be evaluated in care-relevant tasks such as object retrieval, placement, and delivery in simulated and real-world hospital or home-care environments.
Innovation
The novelty of this project lies in integrating zero-shot generalization, cross-domain transfer, and human-like behavior learning into a unified framework for care robotics. Instead of training robots separately for each object or environment, the project emphasizes general manipulation abilities that can extend to unseen care-related objects. It also addresses the practical challenge of limited real-world robot data by transferring knowledge from low-cost, data-rich domains to high-cost target domains. The cross-domain RL component is theoretically grounded through cross-domain Bellman consistency and adaptive hybrid critics. In addition, the project goes beyond task success by considering whether robot behavior resembles human action patterns, using macro action quantization and trajectory optimization to make robots more natural partners in human-centered care environments.
Expected Outcomes
The project is expected to contribute an RL framework for assistive care robots that can generalize, transfer, and collaborate effectively. First, it will provide methods for zero-shot robotic manipulation of diverse care-related objects, supporting applications such as medication handling and delivery of daily necessities. Second, it will advance cross-domain RL methods that reduce reliance on expensive real-world robot training by leveraging simulation and other source domains. Third, it will contribute human-like RL techniques that make robot actions more predictable, interpretable, and suitable for human–robot collaboration. More broadly, the project will offer algorithms, experimental benchmarks, and design principles for deploying learning-based robots in home-care and hospitals ettings, where adaptability, safety, and natural interaction are essential."