ViGAR: The First World-Action Model for Full-Sized Humanoid Robots
智元×北大联合发布 ViGAR:首个面向全尺寸人形机器人的世界—动作模型
In October 2026, Zhuyuan Robotics collaborated with the DA Group at Peking University and CocoMatrix to release ViGAR, the first complete world-action model (WAM) for full-sized humanoid bipedal robots. ViGAR has been validated on the real-world AgiBot A2 platform and achieved strong results on the RoboTwin 2.0 benchmark, ranking third in the real-time leaderboard. The model introduces a hierarchical architecture with a subgoal planner and executor, enabling long-horizon task execution through explicit visual subgoal reasoning. This approach allows the robot to plan and execute complex, multi-step tasks in dynamic environments.
ViGAR represents a significant advancement in world-action models for full-sized humanoid robots, introducing a hierarchical architecture that enables long-horizon task execution through explicit subgoal planning. The model's performance on both simulation and real-world platforms demonstrates its robustness in dynamic, unstructured environments.
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
Reported success rate
74.51%
View original evidence
二、仿真与真机:双线领先 仿真评测:RoboTwin 2.0 Clean2Random 曾登顶榜单 ViGAR 在 RoboTwin 2.0 仿真评测的 Clean2Random 基准上取得 Clean 82.00%、Random 67.02% 的成功率,平均成功率 74.51%,超越最强基线 12.81 个百分点,发布时曾位列榜单第一;该榜单实时更新,截至发稿(2026 年 10 月)位列总榜第三。
Open source S9task_completion_progress
0.69
View original evidence
在 AgiBot A2 全尺寸人形机器人上,ViGAR 对五个长时序组合任务进行真机评测,平均任务完成进度达到 0.69,领先 π0.5(0.60)和使用相同数据配方,但无子目标规划的 Cosmos3-Nano-Policy(0.54)。
Open source S16training_hours
≈900 hours
View original evidence
真机实验采用两条 7 自由度机械臂、两台腕部相机和一台前置相机;研究团队使用约 900 小时的同平台遥操作数据进行中期训练,覆盖 752 项任务、200 余种物体和 10 余种背景,再通过任务示范开展后训练。
Open source S14tasks
752
View original evidence
真机实验采用两条 7 自由度机械臂、两台腕部相机和一台前置相机;研究团队使用约 900 小时的同平台遥操作数据进行中期训练,覆盖 752 项任务、200 余种物体和 10 余种背景,再通过任务示范开展后训练。
Open source S14objects
>200
View original evidence
真机实验采用两条 7 自由度机械臂、两台腕部相机和一台前置相机;研究团队使用约 900 小时的同平台遥操作数据进行中期训练,覆盖 752 项任务、200 余种物体和 10 余种背景,再通过任务示范开展后训练。
Open source S14backgrounds
>10
View original evidence
真机实验采用两条 7 自由度机械臂、两台腕部相机和一台前置相机;研究团队使用约 900 小时的同平台遥操作数据进行中期训练,覆盖 752 项任务、200 余种物体和 10 余种背景,再通过任务示范开展后训练。
Open source S14
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: 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.
它是首个面向全尺寸人形双足机器人的完整世界-动作模型(World-Action Model, WAM),已在真实的全尺寸人形双足机器人 A2 平台上完成多步骤操作验证。
Open source S3
二、仿真与真机:双线领先 仿真评测:RoboTwin 2.0 Clean2Random 曾登顶榜单 ViGAR 在 RoboTwin 2.0 仿真评测的 Clean2Random 基准上取得 Clean 82.00%、Random 67.02% 的成功率,平均成功率 74.51%,超越最强基线 12.81 个百分点,发布时曾位列榜单第一;该榜单实时更新,截至发稿(2026 年 10 月)位列总榜第三。
Open source S9
在物体位置随机化、布局无法从训练数据中记忆的测试场景下,ViGAR 以 67.02% 的成绩在榜单中名列前茅,50 项任务里有 13 项拿到全榜最高分,多于 GigaBrain 的 11 项,PatchWAM、UniWAM 的 10 项,ME-Dex 的 9 项。
Open source S13
在 AgiBot A2 全尺寸人形机器人上,ViGAR 对五个长时序组合任务进行真机评测,平均任务完成进度达到 0.69,领先 π0.5(0.60)和使用相同数据配方,但无子目标规划的 Cosmos3-Nano-Policy(0.54)。
Open source S16
真机实验采用两条 7 自由度机械臂、两台腕部相机和一台前置相机;研究团队使用约 900 小时的同平台遥操作数据进行中期训练,覆盖 752 项任务、200 余种物体和 10 余种背景,再通过任务示范开展后训练。
Open source S14
Source:Chinese robotics — Research news discovery · ithome.com