Darwin Mobile Agent: A Roadmap for Self-Evolution
06:00 · June 23, 2026 · arXiv cs.AI RSS
The goal of artificial intelligence is to create agents capable of general, adaptive behaviour in open-ended environments. Guided by the "Bitter Lesson", we argue that the most effective path toward this goal is to systematically remove human priors and allow intelligence to naturally emerge through interaction with a "Big World" that is orders of magnitude more complex than the agent itself. We propose the mobile Graphical User Interface (GUI) as a practical proxy for such a world and introduce Darwin Mobile Agent, an open-source infrastructure designed as a foundation for autonomous reinforcement learning in this domain. This framework addresses the data-collection bottleneck in real-world mobile interactions by using an asynchronous agent-environment loop across parallel cloud-phone instances. We further propose a conceptual roadmap to systematically remove human priors from three fundamental pillars of a self-evolving agent: task curricula, outcome verification, and memory management. We validate that the Darwin infrastructure provides the stability and scalability required for the first stage of this roadmap: policy optimisation in the GUI domain. This work establishes the practical and theoretical foundation necessary to move toward truly autonomous, self-evolving GUI agents.
Summary
The Darwin Mobile Agent framework provides an open-source infrastructure for training GUI-based agents through reinforcement learning, with the explicit aim of enabling long-term self-evolution. Drawing on the Bitter Lesson, the authors argue that general adaptive behaviour emerges most reliably when human-designed priors are progressively removed and agents learn through sustained interaction with a sufficiently complex environment. They identify the modern mobile GUI as a practical proxy for such a “Big World”: it is partially observable, non-stationary, and contains an effectively unbounded space of composable tasks that evolve independently of any single agent.
To support large-scale data collection, the system replaces conventional ADB emulators with cloud-hosted devices and implements an asynchronous agent-environment loop. A rollout aggregator decouples slow, parallel phone instances from high-throughput policy inference, allowing trajectories to be gathered and verified without blocking model updates. The environment follows a Gymnasium-style interface but deliberately avoids dependence on device XML state, improving both stability and scalability across many concurrent instances.
The paper outlines a conceptual roadmap for removing human priors from three core components of an autonomous agent. Task curricula must shift from hand-crafted sequences to mechanisms that generate problems at the frontier of the agent’s current competence. Outcome verification must move beyond external labels toward self-generated or internally consistent reward signals. Memory management must evolve from static context windows to persistent, queryable agent state that supports non-Markovian reasoning across changing applications. The current release demonstrates the first stage of this roadmap by fine-tuning a UI-TARS policy with reinforcement learning on tasks from Spa-Bench, confirming that the infrastructure sustains stable policy optimisation at scale.
By releasing the full training loop, environment abstractions, and workflow interfaces, the work supplies a concrete foundation for subsequent research on curriculum generation, verification modules, and long-term memory architectures within mobile GUI domains.
Why it matters
This research provides a novel, open-source infrastructure for developing autonomous, self-evolving GUI agents, which is highly actionable for Dutch AI researchers and developers working on reinforcement learning and automation. The focus on removing human priors aligns with advanced AI development goals within the Netherlands' strong technical ecosystem.
