Robot Learning from Limited Demonstrations : From Pile Loading to Scalable Generalization

Abstract

Imitation learning leverages human demonstrations to teach robots complex behaviors, offering a promising alternative to manual programming and trial-and-error. However, practical deployment in unstructured environments faces challenges such as state distribution drift, poor generalization, and multimodal demonstration data. Motivated by an extensive study of autonomous pile loading with heavy machinery across varying seasonal conditions, this thesis identifies the limitations of traditional behavior cloning and lays the groundwork for robust, data-driven solutions that bridge the gap between laboratory success and real-world application.

To extract meaningful learning signals from manually labeled demonstrations without hand-designing dense reward functions, we develop a stage-based visual reward extraction framework that infers task progress directly from observations. For cross-domain adaptation, we propose a sequence-to-sequence framework that decomposes exploration and skill execution, treating hidden environment parameters as latent variables inferred through tactile interaction. Finally, to enable cross-task generalization on offline datasets, we reframe imitation learning as a hierarchical planning problem with subskill discovery, and introduce Vis2Plan, a symbolic-guided visual planning framework that extracts object-centric structure from unlabeled play data and generates visual subgoals executed by a goal-conditioned controller.

This thesis yields three key insights: visual task-progress and terminal signals can be estimated from manually labeled demonstrations, with mixed-season training improving reward-estimator robustness; exploration-conditioned latent-context inference supports adaptation to new domains without test-time retraining; and goal conditioned reformulation unifies multi-objective skill acquisition. The resulting methods reduce complexity and data requirements for real-world robotic deployment, bringing reliable, adaptable imitation learning systems closer to practical use in unstructured environments.