GeneralVLA-2: Giving Robots 'Spatial Memory' to Improve Planning Accuracy
北大联手中科院：给机器人装上“空间记忆”，让它不再靠蒙猜方向_研究_生物体_形状
This research, led by Peking University and the Chinese Academy of Sciences, introduces GeneralVLA-2, a robot planning system that improves geometric perception and memory management. GeoFuse-MV3D enhances 3D shape reconstruction using multi-view geometry, while the governed KnowledgeBank refines memory retrieval with validation and lifecycle management. The system achieves improvements in shape reconstruction metrics and task success rates in both simulation and real-world settings.
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
test_set_size
30
View original evidence
这个测试集来自谷歌扫描物体数据集，包含30个真实家居物品的高精度三维模型，评估协议与MV-SAM3D完全一致：输入相同的五个视角，评估其余视角下的渲染质量。
Open source S12
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
这个测试集来自谷歌扫描物体数据集，包含30个真实家居物品的高精度三维模型，评估协议与MV-SAM3D完全一致：输入相同的五个视角，评估其余视角下的渲染质量。
Open source S12
对比基准方法MV-SAM3D，GeoFuse-MV3D在这四项指标上全部取得了进步：CD下降了2.20%，LPIPS下降了2.02%，PSNR上升了2.36%，SSIM上升了1.03%。
Open source S13
结果来看，KnowledgeBank在所有四个模型骨干上，Terminal-Bench成功率平均比ReasoningBank高出4.53个百分点，SWE-Bench解决率平均高出3.73个百分点，同时平均步数（AS，步数越少说明效率越高）在两个测试集上分别减少了4.95%和5.65%。
Open source S20
Source：Chinese robotics — Research news discovery · sohu.com