CWM: Contrastive World Models for Action Feasibility Learning in Embodied Agent Pipelines

Researchers propose Contrastive World Models (CWM), a new approach for training AI agents to better distinguish between physically feasible and infeasible actions in embodied environments. The method uses contrastive learning with hard negative examples to outperform traditional supervised fine-tuning, achieving 6.76 percentage point improvement in precision and better safety margins under stress conditions.

• →CWM uses contrastive learning with InfoNCE objective to train action feasibility scorers for embodied AI agents.
• →The approach focuses on hard negatives - semantically similar but physically incompatible action candidates.
• →CWM outperforms supervised fine-tuning by +6.76 percentage points on Precision@1 for minimal-edit negatives.
• →Under stress conditions, CWM maintains better safety margins (-2.39 vs -3.96) compared to traditional methods.
• →The research addresses a critical bottleneck in embodied agent pipelines before planning and reasoning occur.