From Failures to Learnable Worlds: Policy-Conditioned Environment Generation for Navigation Agents
WEDGE generates navigation training environments from a policy's validated failure modes, beating random and hard-example baselines.
WEDGE is a framework that uses failure modes of a pretrained navigation policy to generate training environments at each adaptation round. An LLM proposes executable weakness features from contrasting successful and failed rollouts, validates them, and selects the weakness with the largest success-rate gap. Those weaknesses are mapped to controllable environmental factors so environments are built in the difficulty regimes where the policy struggles. Across three agents, weakness-conditioned adaptation outperforms matched random and hard-example methods on ProcTHOR, and gains for Uni-NaVid and Gemini also extend to HM3D.
- WEDGE turns a pretrained navigation policy's failures into targeted training environments.
- An LLM proposes weakness features from successful versus failed rollouts and validates them.
- Weakness-conditioned adaptation beats random and hard-example baselines on ProcTHOR.
- Gains for Uni-NaVid and Gemini also transfer to HM3D after ProcTHOR adaptation.
From the original article
1 min read at huggingface.co
Simulated environments enable scalable training data generation for embodied agents, but more data does not necessarily provide what a particular agent needs to learn. Existing environment-generation pipelines largely optimize generic properties such as realism, diversity, or task coverage, while different pretrained policies can fail for different environmental reasons.
By Yue Zhang, Han Lin, Zun Wang · huggingface.co
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.