Breaking

NVIDIA Unveils World-Action Models for Robots, Releases 14B DreamZero

June 15, 2026 at 15:02 EDT

Robotics
Foundation Models
Research & Papers

On June 15, 2026, NVIDIA unveiled World-Action Models (WAMs), a new approach to physical AI for robots that integrates pretrained scene dynamics (world models) with action generation, aiming to generalize to unseen tasks and environments. The flagship example is DreamZero, a 14B-parameter WAM developed by the company's GEAR lab.

Physical AI · NVIDIA

World Action Models: "Pretrained to Imagine, Fine-Tuned to Act"

NVIDIA's DreamZero — a 14B-parameter video diffusion model — predicts future world states and robot actions together, more than doubling the task performance of conventional Vision-Language-Action models.

14B

parameters, autoregressive video diffusion (DreamZero)

7Hz

real-time closed-loop control rate

~30 min

55 trajectories to adapt to a new robot body

Average task progress — WAM vs VLA

Real-robot experiments on AgiBot data · higher is better

VLA

27.4%

WAM (DreamZero)

62.2%

More than 2× the VLA figure — plus 39.5% on unseen tasks like untying shoelaces and waving.

How a WAM works

Pretrain on video → physics priors

↓

Imagine future world states

↓

Generate & execute robot actions

Built on the Cosmos 3 world model; adopted in next-gen Isaac GR00T 2 (expected by year-end).

Data efficiency

+42%

relative gain on unseen tasks from just 10–20 minutes of video-only demos of other robots or humans.

A dramatic cut from training that once required hundreds of hours.

Well received

Motor execution improves once the model can "dream" the future
Fast human-to-robot transfer, 30-min adaptation
Open-sourced weights & code for self-verification

Open questions

Compute load may stay high after 7Hz tuning
Monocular spatial accuracy: overshoot / undershoot
Some tasks (e.g. opening doors) still hard; awaiting per-robot IDM

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.