world2vec: Mechanistic world models through latent dynamics learning
The authors report faithful long-range dynamics with at least ten times fewer parameters than DreamerV3, TD-MPC2 and LeWM.
World2vec learns from observations and actions without reward information. The paper targets a gap between strong control performance and accurately capturing how an environment behaves. The authors report state-of-the-art control in continuous environments from pixels, alongside accurate system identification across a benchmark of 71 attractors.
ArXiv · AI/CL/LG's note
World2vec learns from observations and actions without reward information. The paper targets a gap between strong control performance and accurately capturing how an environment behaves. The authors report state-of-the-art control in continuous environments from pixels, alongside accurate system identification across a benchmark of 71 attractors.
ArXiv · AI/CL/LG's note
score 5