UNC's WEDGE paper generates training environments targeted at navigation agents' failure modes
UNC-ChapelHill · hf · 2026-10-15
Researchers at UNC Chapel Hill (Yue Zhang, Han Lin, Mohit Bansal, et al.) introduce WEDGE, a weakness-guided navigation learning framework. Existing environment-generation pipelines optimize generic properties like realism and diversity, yet different pretrained policies fail for different reasons — hard-to-see targets, long blind paths, multi-room transitions. WEDGE instead lets the agent's own failure modes guide environment generation each adaptation round: an LLM proposes executable weakness features by contrasting successful and failed rollouts, validates them, and builds targeted training scenarios without a predefined task or skill set.
More from Embodied
- Robotics data tokenization landscape: Khala maps peaq, THOR, BitRobot and Axis — 0xSammy · 2026-10-16
- Perseverance rover logs 28.1 miles on Mars, breaking Opportunity's record in half the time — BrettKrieger12 · 2026-10-16
- Trossen bimanual arms portion food in commercial kitchen with models fine-tuned via Physical Intelligence API — kscottz · 2026-10-16
- Safeworld emerges from stealth with $12M to build 'CI/CD for robots' safety testing — lukas_m_ziegler · 2026-10-16
- Standard Bots launches AI so you can vibe code robot arms from an iPad — Rewkang · 2026-10-16
- Niantic Spatial's places library ships robot-ready Gaussian splat warehouse scenes for policy training — RexDouglass · 2026-10-16