Researchers developed LAWA, a world action model architecture that represents future intentions using compact latent actions rather than generating explicit future observations, employing a discrete tokenizer enhanced through action-free pre-training to create manipulation-centric codebook targets while jointly denoising continuous latent states during inference. On the RoboCasa benchmark, LAWA achieved success rates of 65.6% and 80.8% in few-shot and full-data settings respectively, outperforming Fast-WAM while maintaining comparable performance to Joint-WAM variants with 42.9% lower inference latency.
