CausalWAM Separates Action Causes in World Action Models
Original title: CausalWAM: Learning the Causes of Actions in World Action Models
Sparsity constraint enables causal WAMs to identify action-relevant latent variables, improving closed-loop robotic control efficiency.
The work addresses entanglement of action-relevant and nuisance latent variables in world action models. It proves that under a sparsity constraint action-generating variables form an invertible function of action-relevant latents alone. CausalWAM augments pretrained WAMs with a sparse routing path in the action head while preserving the world-modeling objective. Synthetic and real-robot experiments confirm the theory and show improved closed-loop success rates.
- Sparse routing identifies action-relevant latents
- Theory holds without temporal dependence assumptions
- Requires only sparsity constraint
- Improved closed-loop success on robots
- Validated on synthetic and real-world data
From the original article
1 min read at arxiv.orgWorld action models (WAMs) learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions. The prediction has to be of the whole scene, whereas only a portion of the scene causes actions.
By Juntao Cheng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.