Simulated environments offer scalable training data for embodied agents, but more data is not necessarily what a particular agent needs. The new WEDGE framework argues that a pretrained policy's failures are a better guide: instead of optimizing generic properties like realism or diversity, it generates training environments tailored to the specific weaknesses of the agent being adapted.

WEDGE works by having a language model propose executable "weakness features" from contrasting successful and failed rollouts, then automatically validating them on a diagnostic set and selecting the strongest based on the success-rate gap. Those validated weaknesses are mapped to controllable environmental factors, which are varied to construct environments in the difficulty regimes where the policy struggles.

In experiments with three navigation agents, each policy exhibited a distinct weakness profile and benefited from different training environments. Weakness-conditioned adaptation consistently outperformed matched random and hard-example alternatives on ProcTHOR, and gains for Uni-NaVid and Gemini also transferred to HM3D after ProcTHOR adaptation. Iterative re-diagnosis showed that dominant weaknesses can persist or shift as the policy improves, providing updated guidance for subsequent environment construction.

The results position policy failures not merely as diagnostic signals but as actionable input for building training worlds that target an agent's current limitations, suggesting a move away from generic data scaling toward policy-conditioned curricula.