RoboFoundry: System-as-Policy Evolution for Self-Learning Cross-Embodiment Agents

A foundation model does not act in isolation as an embodied agent. Its behavior is mediated by the context it receives, the experience it retains, the skills it can invoke, and the embodiment interface through which those skills are executed. Moreover, interaction alone does not yield self-improvement unless execution experience is converted into persistent, validated system changes. We therefore propose RoboFoundry, the first embodied agentic framework that formulates this process as Self-Evolving System-as-Policy. RoboFoundry self-diagnoses model-specific capability gaps in decision-making and memory management, then evolves the corresponding system support at task-specific and general scopes. Evolution operates on two complementary system surfaces: a context system that manages active internal context and persistent file-system memory, and a hierarchical skill system that organizes atomic skills, reusable compositions, and failure-conditioned recovery. Both operate over semantic contracts that separate embodiment-invariant decisions from embodiment-specific execution, enabling evolved capabilities to transfer across heterogeneous robots. We evaluate RoboFoundry across diverse embodied settings. On EmbodiedBench, system-level evolution consistently improves decision-making across foundation models, with larger gains for smaller backbones. On RoboMemArena, RoboFoundry achieves state-of-the-art long-horizon memory performance by jointly evolving internal context and external memory use, even against methods assisted by external frontier foundation models. On LIBERO-PRO, it improves robustness under task and environment perturbations through general and task-specific evolution. Real-world experiments further demonstrate zero-shot deployment and online evolution across diverse embodiments and tasks, highlighting the potential of RoboFoundry for autonomous embodied agents.