96 Robots Evolve Together, Success Rate Surpasses Humans, Tsinghua, Zhiyuan Robotics, and Wuwen Xinqiong Propose 'Collective Experience Learning'
96台机器人共同进化，成功率反超人类，清华、智元机器人、无问芯穹联合提出“集体经验学习”
Tsinghua University, Zhiyuan Robotics, and Wuwen Xinqiong jointly propose the Collective Experience Learning (CEL) framework, which uses a method called Failure-Driven Experience Retrieval and Tuning (FERT) to enable robot swarms to share execution experiences and collaboratively improve strategies. In five real-world tasks, FERT increases task success rates to 93.3%-99.2% without relying on expert data. The study also builds a Fault-Tolerant Asynchronous Training System (FATS) to ensure continuous learning even when robots fail.
Source: Chinese robotics — Research news discovery · 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
robot_count
96 robots
Reported success rate
93.3%
View original evidence
在手机包装、插花、叠衣服等五项真实世界操作任务上，研究团队每项任务各进行了 120 次试验，参与评估的机器人均未参与训练数据采集，最终策略的成功率达到 93.3% 至 99.2% 。
Open source S11Reported success rate
99.2%
View original evidence
在手机包装、插花、叠衣服等五项真实世界操作任务上，研究团队每项任务各进行了 120 次试验，参与评估的机器人均未参与训练数据采集，最终策略的成功率达到 93.3% 至 99.2% 。
Open source S11Reported success rate
79.2%
Reported success rate
98.3%
robot_failure_rate
22 minutes
training_trials
120 trials
View original evidence
在手机包装、插花、叠衣服等五项真实世界操作任务上，研究团队每项任务各进行了 120 次试验，参与评估的机器人均未参与训练数据采集，最终策略的成功率达到 93.3% 至 99.2% 。
Open source S11
- paper: Available Artifact link Open source S7
- 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.
在手机包装、插花、叠衣服等五项真实世界操作任务上，研究团队每项任务各进行了 120 次试验，参与评估的机器人均未参与训练数据采集，最终策略的成功率达到 93.3% 至 99.2% 。
Open source S11
96台机器人共同进化，成功率反超人类，清华、智元机器人、无问芯穹联合提出“集体经验学习”
Open source S1
在多阶段手机包装任务中，使用 96 台机器人进行学习后，任务成功率 从 79.2% 提升至 98.3% 。
Open source S13
在 96 台机器人的实际运行中，平均每 22 分钟就有一台出现故障。
Open source S22
二元标签只能说明一条轨迹失败了，却无法定位失败发生的位置。
Open source S19
论文链接：https://arxiv.org/abs/2610.25736 CEL 的算法实现是 失败引导的经验检索微调（FERT） ：利用机器人执行失败时的上下文，从整个集群共享的成功经验中检索相关轨迹片段，并将其中记录的动作作为监督信号来更新策略。
Open source S7
Source：Chinese robotics — Research news discovery · eu.36kr.com