Section
Vision-language-action models
The other recipe: policies built on vision-language models. Pi-0, GR00T and their successors, and the hybrids forming between the two.
A vision-language-action model is a robot policy built on a vision-language model: a system pretrained to describe images and follow language, then adapted to output actions. Physical Intelligence's π0 (Black et al., π0 (2024)) is the example this site uses in the pillar explainer. π0.5 (Physical Intelligence, π0.5 (2025)) is the later open-world generalization paper in that line.
The difference from a world action model is the inherited prior. A VLA inherits knowledge of objects and language. A WAM inherits knowledge of how scenes change. Reuss calls the remaining VLA problem a grounding gap: language-to-action still has to be learned from scarce robot data. That contrast is the reason this section exists on a WAM site — not as a second register of VLAs, which we do not keep, but as coverage of the other recipe and of the hybrids forming between the two. Reuss, NVIDIA Technical Blog (15 June 2026)
Three register rows are already tagged hybrid: Motus, BagelVLA, Cortex 2.0. Motus unifies VLA, world-model, inverse-dynamics and video-generation modes in one Mixture-of-Transformers. BagelVLA interleaves linguistic planning, visual forecasting and actions. Cortex 2.0 adds a world-model planning layer on top of an industrial VLA. Being-H0.7 is filed as latent actions, not hybrid, because that is the register's confirmed paradigm; Reuss still reads it as a VLA–WAM bridge.
Reuss's April 2026 RoboArena snapshot is the public real-world comparison this site cites for the contest: DreamZero at 1,750 Elo, π0.5 at 1,622, with Pi-FAST at 1,592 and Pi-0 at 1,475 in the same figure. Those four numbers are Reuss's leaderboard reading, not a table we re-ran. Reuss, NVIDIA Technical Blog (15 June 2026)
Stories filed here are about that contest: the field map, the hybrids, and the VLA baseline a WAM result has to beat. For the WAM definition itself, use the pillar page and the register.
Models in this section
Stories
Odyssey-3 tests a video world model as a robot backbone
Odyssey-3 is a video world model that Odyssey says can also run robot arms, humanoids and cars. We check the claimed robot results against Physics-IQ Verified.
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.
What is a world action model?
A robot policy that starts from a model pretrained to predict video, then learns to act. Here is how it works, how it differs from a VLA, and what it costs.