wamodels
WAMsVLAs

NVIDIA researcher maps world action models as robotics' second recipe

Moritz Reuss's long post sorts more than a dozen models into three paradigms, and predicts the field will end up with VLA–WAM hybrids.

By World Action Models1 min read

The best single map of world action models so far is a post on NVIDIA's developer blog by Moritz Reuss, published on 15 June 2026. He presents it as a personal map with hypotheses rather than conclusions. It is still the most complete classification of the field we have found, and the register on this site follows its categories.

The argument

Vision-language-action models (VLAs) start from models that understand images and text. Reuss argues they face a "grounding gap": they still have to learn from scarce robot data how language maps onto physical behaviour. World action models (WAMs) start from video models instead, which already know how scenes evolve. He presents this as a different starting point rather than a replacement.

The map

He sorts models on three axes:

  • What they predict. Some generate future video then infer actions (inverse dynamics). Some emit both together (joint prediction). Some use the video backbone's features without generating frames (representation only).
  • How actions are encoded. As tokens, as pictures, or as latent codes learned from video.
  • Architecture. Separate video and action stages, one monolithic transformer, or a mixture-of-transformers, the design also used by Physical Intelligence's π0 family.

The numbers he leans on

  • DreamZero, NVIDIA's 14-billion-parameter WAM, leads RoboArena's April 2026 snapshot at 1,750 Elo, ahead of π0.5 at 1,622.
  • He cites a 3-4x inference slowdown versus π0.5 (Fast-WAM's Fig. 4 gives about 3.2-4.5x for its own Joint and IDM variants).
  • His rough compute estimates put a full WAM at around seven times the training cost of an equivalent VLA path.

The prediction

Reuss expects WAMs to become a core subfield and the eventual winner to be a hybrid of both recipes. He also flags robotics-first foundation models, trained on large wearable and robot datasets rather than borrowed from video or language, as a possible fourth path, limited for now by data access. Benchmarks, he says, remain the field's weakest link.