ViGAR Framework Enables Visual-Goal-Guided Robotic Manipulation for Long-Horizon Tasks
机器人学做饭的时候，到底该先想什么?
Peking University, AgiBot, and CocoMatrix present ViGAR, a framework that decomposes complex robotic tasks into visual subgoal planning and action execution using the Cosmos3-Nano world model. Key innovations include a subgoal look-ahead rule to prevent regression loops, end-effector ROI weighting to focus attention on manipulation regions, and a geometric route for passing pixel-level subgoal images. In simulation, ViGAR achieves an average success rate of 74.51% in random scenarios, demonstrating strong in-context learning capabilities by adjusting task decomposition based solely on global goal images without retraining parameters.
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
22.46%
Reported trials
View original evidence
论文里有个数据很扎心：一个叫Cosmos3-Nano-RoboTwin的对照模型，用和ViGAR完全一样的训练方法，但没有显式的子目标指导,结果在随机场景测试里成功率只有22.46%，而ViGAR同样条件下做到了67.02%。
Open source S11Reported success rate
82%
Reported trials
Reported success rate
67.02%
Reported trials
Reported success rate
74.51%
Reported trials
Reported success rate
82.16%
Reported trials
View original evidence
作者最终选择了几何路线,并给出了一个消融实验来证明这个选择是对的：只用几何路线（生成器单独注入）时,平均成功率是62.17%;如果再叠加语义路线（生成器+推理器同时注入）,Clean场景下的成功率提升到82.16%,但随机场景下的成功率反而从45.42%掉到41.28%。
Open source S17Reported success rate
45.42%
Reported trials
View original evidence
作者最终选择了几何路线,并给出了一个消融实验来证明这个选择是对的：只用几何路线（生成器单独注入）时,平均成功率是62.17%;如果再叠加语义路线（生成器+推理器同时注入）,Clean场景下的成功率提升到82.16%,但随机场景下的成功率反而从45.42%掉到41.28%。
Open source S17Reported success rate
41.28%
Reported trials
View original evidence
作者最终选择了几何路线,并给出了一个消融实验来证明这个选择是对的：只用几何路线（生成器单独注入）时,平均成功率是62.17%;如果再叠加语义路线（生成器+推理器同时注入）,Clean场景下的成功率提升到82.16%,但随机场景下的成功率反而从45.42%掉到41.28%。
Open source S17
- 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.
论文里有个数据很扎心：一个叫Cosmos3-Nano-RoboTwin的对照模型，用和ViGAR完全一样的训练方法，但没有显式的子目标指导,结果在随机场景测试里成功率只有22.46%，而ViGAR同样条件下做到了67.02%。
Open source S11
结果显示,ViGAR在Clean场景下成功率82.00%,在Random场景下成功率67.02%,平均成功率74.51%。
Open source S19
作者最终选择了几何路线,并给出了一个消融实验来证明这个选择是对的：只用几何路线（生成器单独注入）时,平均成功率是62.17%;如果再叠加语义路线（生成器+推理器同时注入）,Clean场景下的成功率提升到82.16%,但随机场景下的成功率反而从45.42%掉到41.28%。
Open source S17
Source：Chinese robotics — Research news discovery · msn.cn