Researchers have introduced ProWAM, a progressive world action model (WAM) designed to improve long-horizon robotic control by jointly predicting actions and an ordered sequence of sparse visual sub-goals. The work, published as a paper on Hugging Face, addresses a key limitation of existing WAMs: generating dense video rollouts is highly inefficient, while predicting only a single future frame neglects how to progress toward a goal.
ProWAM instead provides explicit visual guidance by anchoring action generation to a sequence of sub-goals throughout task execution. According to the paper, this design scales naturally because sub-goal prediction can be learned from large-scale action-free videos, allowing the video backbone to offload complex visual planning from the action policy.
For efficient action generation, ProWAM performs a single video-backbone forward pass to cache sparse sub-goal features, eliminating iterative full-video generation and requiring only lightweight action denoising during replanning.
In evaluations, ProWAM demonstrated superior out-of-distribution robustness. On simulation benchmarks, it set new state-of-the-art results on LIBERO-Plus (85.8%) and randomized RoboTwin (75.7%), outperforming the strongest baseline with relative gains of up to +35.9%. On RoboCasa365, it achieved a 48.1% success rate and 18.2% on the challenging Composite-Unseen split, ranking 4th overall.
Crucially, in zero-shot real-world experiments, ProWAM reached 70.0% success, outperforming the strongest baseline by +15.0 points (from 55.0% to 70.0%, a +27.3% relative gain) in novel scenes. The authors state these results demonstrate the value of progress-indexed visual foresight for closed-loop control. The project page is available at https://sii-ferenas.github.io/ProWAM-page.