Towards deployable autonomous robots

Abstract

Autonomous robots are increasingly expected to operate in industrial yards, warehouses, offices, and outdoor and indoor work sites, where variability in terrain, lighting, weather, and human activity stresses perception, localization, mapping, planning, and control. Achieving deployability under such conditions requires methods that tolerate uncertainty, work within realistic compute/energy and hardware budgets, reduce engineering effort, and make error characteristics explicit. Guided by this perspective, the dissertation advances the algorithmic core of the autonomy stack through four contributions that collectively facilitate real-world deployment.

The first study presents a systematic literature review of autonomous systems for semistructured environments, using mill yards as a representative case. The review maps technologies and sensors across perception, localization and mapping, and path planning and control; discusses feasibility and adoption strategies; highlights implementation challenges; and proposes a conceptual reference autonomy architecture for autonomous navigation and material handling. The second study develops a practical protocol for benchmarking stereo depth sensors and modelling depth errors. Results demonstrate that real-world performance diverges from manufacturer claims, with heavy-tailed error distributions that defy Gaussian assumptions. The study’s approach is inexpensive, reproducible, and offers practical guidance for configuring perception modules.

The third study introduces UDGS-SLAM, which fuses single-image metric depth with 3D Gaussian splatting for dense, photorealistic mapping and pose optimization. UDGSSLAM achieves competitive or superior tracking performance, compared with monocular and RGB-D baselines, as well as high-fidelity view reconstruction. The fourth study presents MoRE, which leverages multiple large language models (LLMs) as concurrent oracles to generate, evaluate, and refine white-box reward functions for reinforcement learning. In 29 Isaac Gym and Dexterity tasks, MoRE matched or surpassed expert, sparse, and single-LLM rewards, achieving top performance on 24 tasks. Ablation studies confirm that diversity, rather than model size alone, drives these gains.

Overall, this dissertation advances the theoretical and practical understanding needed to facilitate real-world deployment of autonomous robots. It contributes: (i) a conceptual reference autonomy architecture for dynamic, semi-structured environments; (ii) a reproducible stereo-depth benchmarking protocol that exposes range-dependent, non-Gaussian error characteristics of depth measurements; (iii) UDGS-SLAM, a monocular photorealistic SLAM system that reduces hardware burden while achieving competitive localization and high-fidelity rendering; and (iv) MoRE, a multi-oracle LLM framework that lowers reward-engineering overhead and delivers strong performance across 29 tasks. Together, these tools and systems reduce engineering effort, enable lean sensing configurations, and improve robustness under real-world noise and variability, and support the wider deployability of autonomous robots.