✦
Summary
UniAction is a released framework for controlling a shared world model with different kinds of actions. It supports force, camera movement, hand pose, and robot pose through one action-conditioning approach, with code and checkpoints available for experimentation.
Key Features
Unified heterogeneous action conditioning.
Force-based world control.
Camera-motion conditioning.
Hand-pose conditioning.
Robot-pose conditioning.
Sparse semantic action elements.
Cardinality-calibrated attention.
Released inference code and checkpoints.
About UniAction
Each action becomes a sparse set of semantic action elements with continuous spatial and temporal coordinates. The model maps those coordinates into a pretrained video model's positional space and uses cardinality-calibrated attention to balance action representations with different element counts.
UniAction helps researchers study compositional action control and generalization across scenes and timing. Its shared representation makes it possible to compare multiple control types within one pipeline rather than maintaining a completely separate conditioning mechanism for every action domain.