VISTA Framework Enables Self-Improving Robots Without Updating Model Weights
机器人自己写代码修自己，成功率从17.5%涨到95%：具身智能有了不动权重的自进化框架
Researchers from Tsinghua University and the National University of Singapore have introduced VISTA, a framework that allows robots to autonomously develop and refine their own skills through code modification. By employing a Meta-agent to iteratively improve Harnesses while keeping underlying model parameters fixed, the system increased task success rates on the LIBERO-Pro benchmark from 17.5% to 95.0%. The approach has been validated in simulation environments, with source code and papers available for public use.
Source: QbitAI — 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
17.5%
Reported success rate
95%
Reported success rate
22.8%
View original evidence
LIBERO-T和LIBERO-S的平均成功率分别从22.8%、12.2%提升至96.6%、93.4%；RoboTwin的50个任务从48.5%提升至74.0%。
Open source S20Reported success rate
12.2%
View original evidence
LIBERO-T和LIBERO-S的平均成功率分别从22.8%、12.2%提升至96.6%、93.4%；RoboTwin的50个任务从48.5%提升至74.0%。
Open source S20Reported success rate
96.6%
View original evidence
LIBERO-T和LIBERO-S的平均成功率分别从22.8%、12.2%提升至96.6%、93.4%；RoboTwin的50个任务从48.5%提升至74.0%。
Open source S20Reported success rate
93.4%
View original evidence
LIBERO-T和LIBERO-S的平均成功率分别从22.8%、12.2%提升至96.6%、93.4%；RoboTwin的50个任务从48.5%提升至74.0%。
Open source S20
- paper: not_reported
- 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.
在底层模型参数保持固定的情况下，VISTA将LIBERO-Pro的整体任务成功率从 17.5% 提升到了 95.0% 。
Open source S8
LIBERO-T和LIBERO-S的平均成功率分别从22.8%、12.2%提升至96.6%、93.4%；RoboTwin的50个任务从48.5%提升至74.0%。
Open source S20
Source：QbitAI — Robotics on WeChat · mp.weixin.qq.com