World action models improve robot control with better video and language training
SWAM: Scaling World Action Models with Cross-Modal Alignment
Computer Vision and Pattern RecognitionRobotics
Summary
Machines can learn to understand videos, language, and actions together to perform tasks better. The authors studied how mixing different kinds of videos and teaching the model with robot actions helps it learn better connections between words, videos, and actions. They found that training on a carefully chosen mix of videos and robot demonstrations helps the model perform well even in new or harder situations. Their approach increased robot success rates in real tasks compared to previous methods. They also provide the data and code so others can build on their work.
What this means in practice
- •For robotics engineers: Improve robot task performance by training models that jointly understand video, language, and robotic actions using diverse video and demonstration data.
- •For video ai developers: Create better systems that align language and video understanding by integrating action learning with large-scale diverse video datasets.
Authors
Yu Zhou, Xinyuan Xia, Xuankun Yang, Yinghao Yang, Wendi Chen, Bowen Liu, Jianhan Ma, Weicheng Duan, Jun Lv, Cewu Lu, Yanwei Li
Abstract
World action models jointly predict video and actions conditioned on language. Existing studies focus primarily on architectures and training mechanisms, typically treating language-video alignment acquired through video pretraining as an established foundation. However, introducing action learning may impair this alignment. Principled guidance on data and supervision design for maintaining language-video alignment during action learning remains limited. To investigate this, we curate a pool of approximately 47,000 hours of general videos, egocentric videos, UMI demonstrations, and robot demonstrations. Building on this, we systematically study how language-video supervision, language-video-action supervision, and the joint use of these two forms affect WAM performance. In particular, we examine how language annotation quality affects pretrained language-video capabilities, and how action supervision from UMI and robot demonstrations extends these capabilities to language-video-action modeling. To better scale the model, we further conduct detailed studies on joint training recipe and model architecture. Under a fixed training budget, mixing egocentric videos and UMI demonstrations improves control generalization. Model architecture affects how effectively these experiences are used, while scaling pretraining under the selected configuration brings further gains. The resulting recipe, post-trained on RoboTwin Clean demonstrations only, achieves success rates of 88.10% and 61.56% under Easy (Clean-to-Clean) and Hard (Clean-to-Random) conditions, respectively. Our model achieves 40.00% SR across three real-world tasks, compared with 23.33% for the strongest baseline, indicating better generalization under unseen conditions. We release the curated data, training code, and model weights.