Event

GAWM Compact World-Action Model Paper Released

1 reports · 1 sources

Yang Liu and other authors released a paper proposing GAWM (Geometry-Aware World Model)—a compact multi-task world-action model that predicts future observations in the latent space of a frozen DINOv3 encoder, with fewer than 100 million parameters and without robot or video pretraining. The authors state that GAWM outperforms billion-parameter-scale WAMs on RoboTwin 2.0 and performs comparably to billion-parameter-scale VLAs on LIBERO; on a dual-arm robot, its action prediction error on held-out demonstrations is lower than that of a fine-tuned billion-parameter-scale VLA, and its inference latency is also notably lower. Ablation experiments show that when training data or training time is limited, the latent world model yields clear performance improvements. These comparison results are all as stated by the authors in the paper.