CRIS-0 pairs a Causality-guided Robot Agent with a Causal World Model, tracking task state to recover from disturbances, run long tasks on its own, and adapt to each user.
Read article →
An embodied world model that predicts motion and geometry before future video — ranking No. 1 on TriWorldBench (66.04) and #1 in the PAI-Bench robot domain.
Read article →
Without updating parameters, open-source models explore new software, verify what they learn, and consolidate causal experience into memory — reaching 78.98% on OSWorld 2.0 and 84.82% on Agents’ Last Exam.
Read article →
Learning a unified latent action interface from visual transitions, so world models can transfer action structure across embodiments instead of overfitting to raw robot commands.
Read article →
Reconstruction-based latent action models can generate convincing video while barely controlled by the action. Debiasing the latent action space first cuts action-following error by over 30%.
Read article →
Contact folds the reachability landscape along a new seam. Interaction-Weighted Resampling restores coverage where it happens — lifting a real air-hockey robot from 5/20 to 12/20.
Read article →
What latent space should a world model learn for control? A single linear projection extracts a compact, task-centric state from frozen visual foundations.
Read article →
Why predictive structure is not causal structure, and how causal world models can give AI a better target for real-world intelligence.
Read article →
A runnable causal analysis agent for discovery, effect estimation, counterfactual reasoning, robustness checks, and inspectable reports.
Read article →
Why the next generation of agents needs memory, imagination, modular control, and causal understanding before it acts.
Read article →
How causal world models can guide exploration, abstract state and action, verify imagined futures, and improve control.
Read article →