Unified Video Action Model
Original research
MODEL PROFILE · World-action research
Unified Video Action
Learns a joint representation for video and robot actions with separate lightweight decoders.
Research organisationLi, Gao, Sadigh and Song
Research approachJoint prediction
Published year2025
BackboneJoint video-action latent representation
Overview
Unified Video Action (UVA) trains coupled video and action predictions, but its decoupled diffusion heads can support action inference without decoding every predicted video frame. The original paper identifies the specific trade-off.
Source record
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.