XPeng and university teams release GroundingPI embodied perception model
XPeng, Peking University, the University of Hong Kong and other research teams released GroundingPI, an approximately 4-billion-parameter embodied perception foundation model. According to the research team, it was trained without joint-action labels on perception-only data covering 11 capability categories and ranked first in an aggregate evaluation of 34 public perception benchmarks. In the team’s comparison, GroundingPI scored 73.68 versus 71.54 for GPT-6 Astra, while the supplied reports contain no independent verification of the ranking or methodology.
GroundingPI provides a released paper, code and model for a perception-first approach to physical AI, with reported results spanning public perception benchmarks and robot simulations.
IT之家 11 月 4 日消息,小鹏集团联合北京大学、香港大学等高校研究团队联合发布具身感知基座模型的研究成果 GroundingPI。
官方表示,该研究提出“感知优先”的技术路线,以约 40 亿参数规模,在 34 项公开感知基准的综合评测中位列第一,并在机器人仿真任务中展现出更高的数据处理效率与更强的场景泛化能力。

▲ GroundingPI 在多个场景和能力上展现了基于视觉基元(构成画面的最小积木块)的感知能力
研究团队将 GroundingPI 与 OpenAI 前沿模型 GPT-6 Astra 在同一套 34 项感知基准上进行对比,按照以上这 11 类能力等权平均,GroundingPI 得分 73.68,Astra 得分 71.54。

#GroundingPI#XPeng#Peking University#University of Hong Kong#GPT-6 Astra#OpenAI