Primary sourceIT之家
View sourceEvent ·Multi-source
小鹏联合北大、港大发布 GroundingPI 具身感知模型,综合得分超 GPT-6 Astra
小鹏联合北京大学、香港大学等团队发布具身感知基座模型 GroundingPI,参数规模约 40 亿。该模型采用纯感知数据训练,不使用带动作标签的机器人数据,覆盖通用检测、指代定位、文档版面等 11 类感知能力。在 34 项公开感知基准上,GroundingPI 等权平均得分 73.68,高于 GPT-6 Astra 的 71.54。仿真实验显示,以它为视觉底座的系统仅用 50% 示教数据即超过对比模型使用 75% 数据的成功率,可降低机器人部署中的人工示教成本。项目已开源,提供网站、arXiv 论文、GitHub 代码和 HuggingFace 模型链接。
Current read
小鹏带着北大港大做的具身感知模型,40亿参数在34项基准上跑赢 GPT-6 Astra,训练还不用机器人动作数据,代码模型都开源了。
11 sources62° AI heatLast updated 2026/11/4 09:40:04
Evidence
11 sourcesRelated source 1elvis
View sourceRelated source 2Artificial Analysis
View sourceRelated source 3rohanpaul_ai
View sourceRelated source 4marktechpost
View sourceRelated source 5kimmonismus
View sourceRelated source 6Thomas Wolf
View sourceRelated source 7Gorden Sun
View sourceRelated source 8歸藏(guizang.ai)
View sourceRelated source 9shao__meng
View sourceRelated source 10The Rundown AI
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.