Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning
Original research
MODEL PROFILE · World-action research
Cosmos Policy
Encodes robot actions as latent video frames while predicting future states and actions.
Research organisationNVIDIA research
Research approachJoint prediction
Published year2026
BackboneCosmos-Predict2-2B
Overview
Cosmos Policy adapts the Cosmos-Predict2-2B video backbone. It represents actions, predicted future states and values as latent frames inside the video diffusion process; NVIDIA places it in the joint-prediction category.
Source record
- Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning
- NVIDIA WAM technical synthesis (Moritz Reuss, 15 Jun 2026)
Core research identity and broad method were source-checked as of 2026-11-03. Independent benchmarking, commercial availability, weights and licensing have not been assessed.