FLEX-WAM is a joint model of video states and actions trained with flow matching. It uses block-causal attention with a block size that can be changed at deployment time, so the same checkpoint can predict one frame at a time or several frames in parallel while remaining KV-cacheable. Training randomizes block size and balances state and action gradient contributions, with a forward-dynamics elasticity term that keeps predicted futures responsive to commanded actions. The result is stable open-loop rollouts for thousands of steps and better latency than fixed-horizon alternatives. In MCTS, the model serves as both action proposer and simulator, solving long-horizon PushT and all five OGBench Puzzle-4x4 tasks purely in imagination. On a bimanual OpenArm robot, one checkpoint acts as both play policy and expected-outcome predictor, letting the system spot model-reality mismatches in real time.
Flexible block-causal models sustain efficient long-horizon action-conditioned rollouts
FLEX-WAM uses one checkpoint to predict futures, generate actions and support planning across variable context and prediction lengths.
Center for Robotics and Embodied Intelligence (CREO), New York University · Courant Institute of Mathematical Sciences · Center for Data Science, New York University · Ecole normale supérieure - PSL · NVIDIA
In one line
A block-causal world-action model enables real-time, stable long-horizon imagination and planning, solving benchmark tasks entirely in imagination.
Before you read this
Three ideas this paper builds onThe authors trained a flow-matching model over video states and actions, assigning independent noise levels to each modality and timestep.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.