ZTE Unveils EmbodiedMind Framework for Long-Horizon Robot Planning
Overview
ZTE Communications has published a paper titled 'EmbodiedMind: Adaptive Data Curation and Prefix-Tree Reinforcement Learning for Efficient Embodied Intelligence,' which introduces a three-stage training framework (RSFT, IR-GRPO, Trie-GRPO) to improve long-horizon planning in embodied intelligence. The 8B model achieved an average score of 70.02% across 18 benchmarks using 522k training samples. The 180B Nebula-EmbodiedBrain model ranked first domestically on the SuperCLUE EmbodiedCLUE-VLA leaderboard in September 2026 with a safety score of 97.92. Performance was validated through simulation and physical tests on AgiBot A2 robots.
Generated from attributed reports · Updated 1 hours ago
Event evidence and corrections
1 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · robots: 40000 other · reported trials
“据中国电子学会数据，2026 年上半年国内人形机器人出货量已超 4 万台，全球占比高达 97%；海外科技巨头亦在加速入场：OpenAI 首席执行官已确认公司将自主研发人形机器人，并直言机器人外观 “远不如真正弄清楚如何让机器人大脑运转重要”，计划 2027 年推出演示样机，核心攻关方向正是驱动机器人运转的智能系统。”
Exact source · revision 1jiqizhixin
Reported quantity · Reported accuracy: 30.31 percent · Basis not reported
“在这一背景下，中兴通讯发布论文《EmbodiedMind: Adaptive Data Curation and Prefix-Tree Reinforcement Learning for Efficient Embodied Intelligence》，以 8B 模型在 18 项具身大脑基准上取得 70.02% 的平均成绩，长程规划准确率较现有最优方法提升 30.31%，相关论文已投稿 ICRA 2027。”
Exact source · revision 1jiqizhixin
Reported quantity · samples: 522000 other · Basis not reported
“图 3 Trie-GRPO 自动迭代训练框架（策略与奖励模型联合训练） 从 18 项基准到物理世界，用数据证明 EmbodiedMind-8B 仅使用 52.2 万训练样本，便在 18 项具身大脑基准上取得 70.02% 的平均成绩，综合能力超越同量级开源模型。”
Exact source · revision 1jiqizhixin
Report timeline
Follow attributed reports and material updates.
- Machine Heart — Robotics on WeChatSignalZTE Unveils EmbodiedMind Framework for Long-Horizon Robot Planning
ZTE Communications has published a paper titled 'EmbodiedMind: Adaptive Data Curation and Prefix-Tree Reinforcement Learning for Efficient Embodied Intelligence,' which introduces a three-stage training framework (RSFT, IR-GRPO, Trie-GRPO) to improve long-horizon planning in embodied intelligence. The 8B model achieved an average score of 70.02% across 18 benchmarks using 522k training samples. The 180B Nebula-EmbodiedBrain model ranked first domestically on the SuperCLUE EmbodiedCLUE-VLA leaderboard in September 2026 with a safety score of 97.92. Performance was validated through simulation and physical tests on AgiBot A2 robots.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).