10 Recent Embodied AI Papers from OpenAI, Alibaba DAMO Academy, and Others
OpenAI/阿里达摩院等都在卷，10篇具身智能论文汇总，从统一具身大脑到机群自我进化
This article summarizes ten recent research papers in embodied AI and robotics from institutions including OpenAI and Alibaba DAMO Academy. The selected works cover topics such as unified embodied brains, recursive self-improvement in physical AI, dexterous bimanual tracking control, cross-embodiment world models for locomotion and manipulation, human-interface VLA agents for open-world games, collective experience learning across robot fleets, lightweight frameworks for robot coding agents, evaluation of large language models in embodied tasks, holistic world intelligence from human ego-experience, and egocentric representations from human experience. The collection aims to provide new perspectives for building next-generation robots with enhanced understanding, action capabilities, and continuous evolution.
Source: Chinese robotics — Robot learning research · 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
Reported success rate
51.5%
Reported trials
View original evidence
评估结果显示，RynnUnify 在统一多模态模型中取得最高的具身理解平均分，Rynn-VP 整体得分较 Cosmos 3 Nano 提升 9.75 分，智能体扩展在 Rynn-GEU 上较直接推理提升 6.0 分，世界-动作扩展在 20 个 RoboCasa365 复合任务上达到 51.5% 成功率、三项真实机器人任务平均成功率为 46.7%。
Open source S9Reported success rate
46.7%
Reported trials
View original evidence
评估结果显示，RynnUnify 在统一多模态模型中取得最高的具身理解平均分，Rynn-VP 整体得分较 Cosmos 3 Nano 提升 9.75 分，智能体扩展在 Rynn-GEU 上较直接推理提升 6.0 分，世界-动作扩展在 20 个 RoboCasa365 复合任务上达到 51.5% 成功率、三项真实机器人任务平均成功率为 46.7%。
Open source S9Reported success rate
86.79%
Reported trials
View original evidence
在无微调条件下，单一检查点在训练物体未见轨迹上达到 86.79% 成功率、留出物体上达到 57.32%，分别较 SOTA 基线提升 77.36 和 45.73 个百分点。
Open source S14Reported success rate
57.32%
Reported trials
View original evidence
在无微调条件下，单一检查点在训练物体未见轨迹上达到 86.79% 成功率、留出物体上达到 57.32%，分别较 SOTA 基线提升 77.36 和 45.73 个百分点。
Open source S14Reported success rate
37.4%
Reported trials
View original evidence
HiSTEVE 在 MineStudio Bench 上达到 85.0% 的原子任务得分、在 OpenHA 上平均成功率为 37.4%，并能像人类玩家一样从零完成锻造铁镐等长时程任务。
Open source S19Reported success rate
99.2%
Reported trials
View original evidence
配合容错的异步训练系统（FATS），在 5 个真实操作任务上达到 93.3%–99.2% 成功率（每任务 120 次试测、机器人均未参与训练数据采集）。
Open source S22
- 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.
评估结果显示，RynnUnify 在统一多模态模型中取得最高的具身理解平均分，Rynn-VP 整体得分较 Cosmos 3 Nano 提升 9.75 分，智能体扩展在 Rynn-GEU 上较直接推理提升 6.0 分，世界-动作扩展在 20 个 RoboCasa365 复合任务上达到 51.5% 成功率、三项真实机器人任务平均成功率为 46.7%。
Open source S9
在无微调条件下，单一检查点在训练物体未见轨迹上达到 86.79% 成功率、留出物体上达到 57.32%，分别较 SOTA 基线提升 77.36 和 45.73 个百分点。
Open source S14
HiSTEVE 在 MineStudio Bench 上达到 85.0% 的原子任务得分、在 OpenHA 上平均成功率为 37.4%，并能像人类玩家一样从零完成锻造铁镐等长时程任务。
Open source S19
配合容错的异步训练系统（FATS），在 5 个真实操作任务上达到 93.3%–99.2% 成功率（每任务 120 次试测、机器人均未参与训练数据采集）。
Open source S22
Source：Chinese robotics — Robot learning research · news.qq.com