World Action Model
InternW0-Δ
An Embodied World Model Bridging Predictive Dynamics and Actions
Jointly learning visual dynamics and robot actions,
with Causal Imprint capturing action-relevant future changes.
Read the paper and explore code and models.
Demonstrations
Real-world
robot manipulation.
Explore demonstrations with grippers
and dexterous hands.
Qualitative demonstrations. Playback-speed annotations, where shown, are part of the supplied footage.
Data curation
Filtering the
training data.
Examples of filtered or masked segments,
grouped by the reason for exclusion.
Robot and human demonstrations are curated and aligned for pretraining. The examples below illustrate four types of data filtering, including segment removal and action-target masking.
Method
World Action Model
with directed attention.
A pretrained video expert and an action expert,
with separate parameters and masked interaction.
- Sparse Memory Context (SMC)
- Anchor, recent, and current frames preserve episode context and recent execution history without a dense video history.
- Task-conditioned scene semantics
- A frozen VLM encodes the instruction together with current observations, providing scene-grounded context for the action expert.
Directed attention
Video and action experts exchange information through masked joint attention while retaining their own cross-attention and feed-forward layers. Causal Imprint tokens (Δ) attend to R, C, and Δ; action tokens attend to A, R, C, Δ, and Act. Neither can attend to future video tokens F.
Causal Imprint (CI)
Learnable queries capture action-relevant future changes from recent and current observations. Latent-difference supervision and future-feature alignment shape these representations, which the action expert uses alongside observed visual features.
4D-aware representation distillation
A frozen Track4World teacher provides offline-cached geometry and motion descriptors. A query decoder aggregates clean A/R/C features, and an auxiliary loss distills these priors into the video expert during training.
Experiments
Results across
benchmark settings.
LIBERO-Plus, RoboTwin,
RoboDojo, and EBench.
All axes start at zero. RoboDojo uses a 0–30% range; the other panels use 0–100%.
Open research
Paper, code,
and models.
Explore the paper, code repository, and model collection.