Stamatopoulou, Maria; (2025) Leveraging Demonstrations to Advance Quadrupedal Autonomous Locomotion and Planning. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Stamatopoulou_10218183_Thesis_sigs_removed.pdf Download (47MB) | Preview

Abstract

Advancing autonomy in quadrupedal robots requires progress in both locomotion and planning, ensuring efficient movement while adapting to real-world variability. This dissertation addresses this challenge by leveraging demonstrations as a foundation for skill learning and path planning, focusing on two central problems: how quadrupeds can efficiently acquire and integrate diverse locomotion gaits, and how they can generate feasible and generalizable trajectories. The first contribution, eGAIT, introduces a hierarchical reinforcement learning framework for multi-gait integration, where demonstration-optimized gait policies are fused through an eDQN-based selector with auxiliary guidance. This enables smooth, efficient transitions across velocity-dependent gaits, addressing the challenge of orchestrating multiple skills for energy-efficient locomotion. However, orchestration presupposes access to a diverse skill library, which is costly to construct at scale. To address this, SDS proposes a framework for learning locomotion gaits directly from single video demonstrations. By converting raw videos into executable reward functions via vision–language models, SDS removes the need for handcrafted rewards, motion capture, or manually designed metrics, enabling scalable skill acquisition from natural exemplars. With locomotion capabilities in place, quadrupeds must also reason about where to go. DiPPeR reframes global path planning as an image-conditioned diffusion process trained on map–trajectory demonstrations, producing collision-free global paths more efficiently than classical search-based planners while generalizing to unseen maps. Recognizing that static global plans fail in dynamic environments, DiPPeST extends this framework to local replanning and execution. By combining optical-flow waypoint tracking with ROI-based goal correction, DiPPeST achieves zero-shot online adaptation without retraining, demonstrating robustness in dynamic and partially observable settings. Together, these contributions establish demonstrations as a unifying substrate for quadrupedal autonomy: reducing manual engineering, enabling reusable multi-skill locomotion, and providing demonstration-derived priors for efficient, adaptive planning. This work advances quadrupeds toward scalable, efficient, and reliable autonomy in real-world environments.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item