RoboFoundry: System-as-Policy Evolution for Self-Learning Embodied Agents

RoboFoundry. System-as-Policy Evolution for Self-Learning Embodied Agents. By Jingsong Liang and seventeen others, from Nanyang Technological University, Beihang University, and three other institutions. Published on the twenty-sixth of September. No code repository is linked to it.

The opening position is that a foundation model should not act alone as an embodied agent. Around it sits a stack: memory, context, skills, action interfaces. Existing methods tend to optimise one of those parts at a time, rather than treating the supporting system as a single policy. There is a second gap. Interaction by itself does not make an agent better. Improvement only happens when execution experience is turned into changes to the system that persist and have been validated. An agent can act many times and end up exactly where it started, if nothing it went through is written back into the system that shapes its next attempt.

RoboFoundry, described as the first embodied agentic framework of its kind, formulates that process as a Self-Evolving System-as-Policy. The whole supporting system is what gets improved. It diagnoses gaps in capability, both in decision-making and in how memory is managed. It converts execution traces into validated updates specific to a task. And improvements that keep recurring are promoted into the general system. There are two tiers, then. A fix begins life as something validated for a particular task, and only when it keeps proving useful does it graduate into the system that every task shares.

Evolution happens on two complementary surfaces. One is a context system, which manages the active internal context and a persistent memory kept in a file system. The other is a hierarchical skill system, which organises atomic skills, reusable compositions of them, and recovery routines conditioned on failure. Underneath both lies a shared semantic interface that separates decisions that do not depend on the body from execution that does. That separation is what lets evolved capabilities transfer across robots of different kinds.

On EmbodiedBench, RoboFoundry reports state-of-the-art performance, notably improving GPT-5.5 by twenty-seven point eight percent. It brings Qwen3.7-Plus to near parity with GPT-5.5, seventy point three percent against seventy-two point seven, which the authors take as consistent gains across foundation models. For long-horizon memory, it beats every baseline on RoboMemArena by at least thirty-nine percent, including methods assisted by external foundation models.

On LIBERO-PRO it outperforms Cap-Agent0 by between two hundred forty-three point eight and six hundred seventy-nine point seven percent, and that holds across every perturbation type rather than a favourable few. Off the simulator, in real-world deployments, it shows zero-shot transfer and online evolution across robots and tasks. Those last two go together: the system carries what it has learned to a new robot, and keeps evolving once it is there.

Treat the scaffolding as the thing being learned, and the question shifts from which model to use to how the system around it grows. RoboFoundry's answer is a loop of diagnosis, validated updates, and promotion, split between context and skills, and held apart from any one robot's body. The authors present it as a step toward fully autonomous embodied agents.