Embodied AI systems face three fundamental challenges in real-world deployment: handling diverse tasks across multiple domains, adapting to unpredictable environmental changes over time, and transferring learned policies across different physical plat...
Embodied AI systems face three fundamental challenges in real-world deployment: handling diverse tasks across multiple domains, adapting to unpredictable environmental changes over time, and transferring learned policies across different physical platforms with varying embodiments. These characteristics define open-domain environments where traditional approaches fail due to catastrophic forgetting in multi-task learning, limited real-time adaptation capabilities, and poor generalization across embodiments.
This dissertation presents a unified framework for skill-centric hierarchical meta-learning that addresses these challenges through a novel two-level architecture. Our approach leverages skills, defined as semantically meaningful action sequences executed over specific time durations, as a fundamental abstraction. The framework decomposes decision-making into a high-level task planner that composes skill sequences to execute complex instructions and a low-level controller that reconstructs given skills into embodiment-specific action sequences suited for the current environment.
For multi-task environments, we develop skill-regularized task decomposition methods that extract and learn high-quality skills from heterogeneous datasets of varying quality. By leveraging reward information to constrain the skill distribution toward high-reward behaviors, our approach achieves performance improvements of 8.67 to 17.67 percent over competitive baselines across diverse dataset configurations. For dynamic environments, we introduce two complementary mechanisms: exploratory retrieval-augmented planning that maintains and queries temporal knowledge graphs for continual adaptation, and temporality-based decomposition embedding that separately captures time-variant and time-invariant features while enhancing training stability. These methods demonstrate 16.45 percent improvement in task success rate, with the advantage becoming more pronounced as environmental non-stationarity increases. For cross-embodiment adaptation, we propose world model implanting that retrieves and integrates domain-specific world models through a two-stage alignment process, and in-context skill diffusion that disentangles skills from embodiment-specific information. These frameworks achieve 20.41 percent improvement in zero-shot scenarios and 26.58 percent improvement in few-shot adaptation settings.
Comprehensive experiments across diverse simulation environments including Meta-world for robotic manipulation, VirtualHome and ALFRED for household tasks, CARLA for autonomous driving, and AirSim for drone navigation validate the effectiveness of our approaches. Our methods consistently outperform state-of-the-art baselines across all three challenge domains, providing a foundation for building robust and adaptable embodied agents capable of operating effectively in real-world deployment scenarios.