Xinzhiyuan · WeChatJun 2, 2026, 12:00 UTC+8

CAS Opens MobileGym, a Browser-Based Agent Training Environment for Mobile Apps

浏览器秒变手机！中科院开源Agent训练场，微信、原神都能跑

CASIA released MobileGym, a browser-based Android simulation environment covering 28 apps, with about 400MB per instance, 3-second cold start, JSON state snapshots, and programmatic task verification for mobile-agent training and evaluation.

Why it matters: MobileGym is practical open-source infrastructure for agent training and evaluation, with enough concrete numbers and mechanisms to pass HKR-H/K/R. It fits the 78–84 quality band, below major lab model-release weight.

Read the original ↗Export Markdown

Take

MobileGym is useful because it attacks the dirty-environment problem before chasing leaderboard theater. The concrete hooks are strong: 28 simulated apps, about 400MB per instance, 3-second cold starts, JSON state snapshots, and a 256-task evaluation finished in 6 minutes. With GRPO, Qwen3-VL-4B moved from 9.4% to 22.2% test success.

The part I buy is programmatic verification, not the phone-in-browser demo. In 118 real-phone trajectories, a VLM judge missed 12 cases; GPT-5.4 still had a 10.2% error rate, just on different tasks. AndroidWorld already had verifiable tasks, but it did not reach WeChat or Alipay-style daily apps. MobileGym dodges accounts, risk controls, reset pain, and parallel rollout limits through simulation. The bill comes later: app realism has to be earned task by task.

Source: Xinzhiyuan · WeChat · mp.weixin.qq.com
Published: Jun 2, 2026, 12:00 UTC+8 · Jun 2, 2026