Event
JAM framework unifies VLA and WAM action models
2 reports · 2 sources
The authors propose the Just Action Model (JAM) framework, which decomposes the differences between vision-language-action models (VLA) and world-action models (WAM) into three separable choices: backbone initialization, future observation objectives, and whether to retain the lookahead loss at inference time. JAM uses a pretrained vision backbone plus an action expert, trained with an action loss and a lookahead loss; the lookahead loss makes current observation features predict frozen encodings of recorded futures, and the prediction head is discarded after training, so neither training nor inference involves visual generation, and the backbone can be chosen freely. The authors report that the JAM policy outperforms the corresponding FastWAM and ImageWAM on LIBERO and LIBERO-Plus, and achieves similar or slightly higher success rates on RoboTwin; when cached lookahead targets are used, its estimated online training FLOPs per sample per update are 73% lower than those of FastWAM.
Report timeline
- JAM: A Unified Action Model Framework for VLA and WAM
The authors propose the Just Action Model (JAM) framework, which breaks down the differences between VLA and WAM into three separable choices: backbone initialization, the loss objective on future observations, and whether that loss mechanism is retained at inference. JAM unifies both into a structure of a pretrained visual backbone plus an action expert, trained with an action loss and a foresight loss.
arXiv cs.RO ·
- JAM: A Unified Action Model Framework Decoupling Backbone, Objective, and Inference Mechanism
The study proposes the Just Action Model (JAM) framework, which breaks down the differences between VLA and WAM into three separable choices: backbone initialization, the future-observation objective, and whether that loss mechanism is retained at inference. It uses a pretrained visual backbone plus an action expert, trained with an action loss and a foresight loss.
HuggingFace Daily Papers ·