A Reproducible, Resource-Efficient Benchmark for Foundation-Model-Inspired Manipulation Learning with LeRobot and Open X-Embodiment

This paper presents a reproducible, resource-efficient benchmark for foundation-model-inspired manipulation learning using LeRobot and the Open X-Embodiment dataset. Four policy architectures are implemented and evaluated under a standardised training and evaluation pipeline: two MLP baselines, a simplified vision-language-action (VLA) model, and a simplified diffusion-inspired policy. The experiments focus on offline imitation learning using real-robot manipulation data and are designed to be reproducible on standard academic CPU hardware. The work provides a modular and low-cost framework for evaluating foundation-model-inspired approaches to robotic manipulation and supports future extensions toward vision, language, touch sensing, and real-robot applications.