MM-ABC Enables Robots to Move and Manipulate Simultaneously
让机器人学会"边走边干活"：MM-ABC如何解决移动操作的两大难题
Researchers from HKUST (Guangzhou) and Tsinghua University propose MM-ABC, a foundation model for general mobile manipulation. The system integrates multi-layered visual perception, future geometric prediction supervision, and coordinated arm-base control to enable robots to move and manipulate objects simultaneously. Trained on over 5000 hours of data from diverse robotic morphologies, MM-ABC achieves a 44.71% success rate on the EBench benchmark, outperforming prior methods like π0.5 by 3.3 percentage points.
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
training hours
>5,000 hours
View original evidence
研究团队构建了一个规模相当可观的预训练数据集，总计超过5000小时，来自12个数据集、51个子集、17种不同的机器人形态，囊括了超过40万条演示轨迹和6万5千多条独特的自然语言指令。
Open source S24demonstrations
>400,000 episodes
View original evidence
研究团队构建了一个规模相当可观的预训练数据集，总计超过5000小时，来自12个数据集、51个子集、17种不同的机器人形态，囊括了超过40万条演示轨迹和6万5千多条独特的自然语言指令。
Open source S24Reported success rate
32.8%
Reported trials
View original evidence
如果不采用清晰动作预测，而是坚持用速度预测，论文的消融实验显示，在RoboCasa365基准测试的复合任务上，成功率会从32.8%直接掉到29.2%，掉了3.6个百分点。
Open source S19Reported success rate
29.2%
Reported trials
View original evidence
如果不采用清晰动作预测，而是坚持用速度预测，论文的消融实验显示，在RoboCasa365基准测试的复合任务上，成功率会从32.8%直接掉到29.2%，掉了3.6个百分点。
Open source S19
- paper: 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.
研究团队构建了一个规模相当可观的预训练数据集，总计超过5000小时，来自12个数据集、51个子集、17种不同的机器人形态，囊括了超过40万条演示轨迹和6万5千多条独特的自然语言指令。
Open source S24
如果不采用清晰动作预测，而是坚持用速度预测，论文的消融实验显示，在RoboCasa365基准测试的复合任务上，成功率会从32.8%直接掉到29.2%，掉了3.6个百分点。
Open source S19
用五步采样（也就是机器人实际部署时用的采样步数）生成动作时，清晰动作预测的任务分配误差比速度预测低了14%。
Open source S17
Source：Chinese robotics — Research news discovery · sohu.com