Embodied Intelligence Begins to Evolve Systems: RoboFoundry Enables Robot Autonomous Evolution
具身智能开始卷系统了：RoboFoundry让机器人自主进化
RoboFoundry is a new framework for embodied agents that treats the entire system as a self-evolving policy. By freezing the base model and allowing the surrounding system to evolve, it achieves significant improvements in task success rates across benchmarks like EmbodiedBench, RoboMemArena, and LIBERO-PRO. The system enables zero-shot transfer and online evolution on real robots, with results showing substantial performance gains compared to baselines.
RoboFoundry introduces a novel framework for self-evolving embodied agents by treating the entire system as a policy. It enables continuous improvement through task execution, reflection, and adaptation, with results showing significant performance gains across multiple benchmarks and real-world robots.
Source: Machine Heart — Robotics on WeChat · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
56.9%
Reported success rate
78%
Reported success rate
39%
- paper: Available Artifact link Open source S11
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
在真实机器人上，展示了跨机器人、跨任务的零样本迁移，以及部署后的在线进化。
Open source S10
在 EmbodiedBench 上，直接使用 GPT-5.5 的平均成功率是 56.9%；模型参数一个不改，只让它周围的系统持续进化，这个数字变成了 72.7%。
Open source S3
图 3 是一次进化过程示例，在 EB-Habitat 的空间关系子集上，基座模型为 GPT-5.5，共经过六次候选评估： 前三次修改被提交，held-in 成功率从 21.4% 升至 50.0%； 接下来两个候选只达到 42.9%，未被采纳； 最后一个候选把成功率推到 78.0%，并通过了 held-out 检查，被标记为晋升进通用系统。
Open source S16
进化后的 GPT-5.5（72.7%）已经高于未经进化的 GPT-6 Astra（71.9%）。
Open source S22
GPT-5.5 相对提升 27.8% 。
Open source S21
GPT-6 Astra 相对提升 8.5%。
Open source S23
消融实验显示，通用进化是关键：以 Qwen3.7-Plus 为基座时，完整版比只保留任务级修改的 Lite 版相对高出 11.2%。
Open source S24
实验中，RoboFoundry 在 EmbodiedBench 上让五个基座模型全部取得提升；在长程记忆基准 RoboMemArena 上，任务成功率比所有基线至少高出 39.0%；在 LIBERO-PRO 的扰动测试中，比 CaP-Agent0 高出 243.8–679.7%。
Open source S9
论文标题：RoboFoundry: System-as-Policy Evolution for Self-Learning Embodied Agents 论文链接：https://arxiv.org/abs/2609.32862 项目主页：https://jingsongliang.com/robofoundry 图 1：RoboFoundry 总览。
Open source S11
Source：Machine Heart — Robotics on WeChat · mp.weixin.qq.com