UniWAM Unifies Physical Reasoning, World Generation, and Action Prediction for Robotic Manipulation
机器人学会了“看图说话”再动手：UniWAM如何让AI既懂世界又会干活
Researchers from HKUST (Guangzhou) and OLA Dimensions present UniWAM, a unified system integrating physical reasoning, world generation, and action prediction to address limitations in existing Vision-Language-Action (VLA) and World-Action (WAM) models. The model employs a Mixture-of-Transformers architecture with joint multimodal attention and physical language supervision. On the LIBERO benchmark, UniWAM achieves an average success rate of 99.2%, outperforming prior methods like Xiaomi-Robotics-0 (98.7%). In out-of-distribution tests on LIBERO-Plus, it reaches 92.6% overall success. Real-world experiments on an AgileX Piper platform show a 67.5% average success rate across four instruction-following tasks.
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
total training hours
≈10,013 hours
View original evidence
这三类数据加起来总时长约10013小时,机器人数据约4958小时(来自AgiBot World、Bridge、Droid、Fractal、Robocoin、InternData-A1等六个来源),人类数据约5072小时(来自EgoDex、EgoVerse、VITRA)。
Open source S12robot data hours
≈4,958 hours
View original evidence
这三类数据加起来总时长约10013小时,机器人数据约4958小时(来自AgiBot World、Bridge、Droid、Fractal、Robocoin、InternData-A1等六个来源),人类数据约5072小时(来自EgoDex、EgoVerse、VITRA)。
Open source S12human data hours
≈5,072 hours
View original evidence
这三类数据加起来总时长约10013小时,机器人数据约4958小时(来自AgiBot World、Bridge、Droid、Fractal、Robocoin、InternData-A1等六个来源),人类数据约5072小时(来自EgoDex、EgoVerse、VITRA)。
Open source S12noise retention ratio range
0.5–1
View original evidence
于是论文的做法是,在训练时故意往未来帧的视觉特征里掺入一定比例的高斯噪声(50%概率触发,噪声保留比例在0.5到1.0之间随机采样),逼着动作预测器学会从一个"模糊、不完美"的未来画面里,依然能提炼出控制相关的关键信息,而不是死死依赖一张清晰但可能是错的未来画面。
Open source S24noise retention ratio range upper
0.5–1
View original evidence
于是论文的做法是,在训练时故意往未来帧的视觉特征里掺入一定比例的高斯噪声(50%概率触发,噪声保留比例在0.5到1.0之间随机采样),逼着动作预测器学会从一个"模糊、不完美"的未来画面里,依然能提炼出控制相关的关键信息,而不是死死依赖一张清晰但可能是错的未来画面。
Open source S24
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
这三类数据加起来总时长约10013小时,机器人数据约4958小时(来自AgiBot World、Bridge、Droid、Fractal、Robocoin、InternData-A1等六个来源),人类数据约5072小时(来自EgoDex、EgoVerse、VITRA)。
Open source S12
于是论文的做法是,在训练时故意往未来帧的视觉特征里掺入一定比例的高斯噪声(50%概率触发,噪声保留比例在0.5到1.0之间随机采样),逼着动作预测器学会从一个"模糊、不完美"的未来画面里,依然能提炼出控制相关的关键信息,而不是死死依赖一张清晰但可能是错的未来画面。
Open source S24
Source：Chinese robotics — Research news discovery · msn.cn