LC-IL: language critiques turn suboptimal demos into robot skills, +7 pts over baselines
shaohua0116 · x · 2026-11-13
Presented at CoRL 2026 workshops, Language-Critique Imitation Learning (LC-IL) teaches robots from imperfect demos by labeling what to do, what went wrong, and how to correct it—via structured language critiques (task progress, action quality, movement guidance) generated per frame. A fine-tuned SmolLM2-135M captioner and a differentiable LC-loss plug into BC and diffusion policies (LC-BC/LC-DP) with zero test-time overhead. Across 7 continuous-control tasks, LC-BC averages 67.7 and LC-DP 72.6, beating BC, DPO, ILWD, and DemIDIE by 7+ points; it hits 87% with only 33% expert purity in mixed-quality data and 92.6% with just 10% expert demos.
More from Embodied
- Video shows an autonomous parking robot handling the full parking job on its own — DavidWells · 2026-11-13
- BoSS benchmark released: 180K+ stellarators, 3TB of simulation data for generative models — ZhongingAlong · 2026-11-13
- Fly brain hooked up to a synth learns to sing Bad Apple — Bedrovelsen · 2026-11-13
- CoRL takeaways: SuperMimic steals the show, dexterity and tactile sensing improve, human data still a bottleneck — KyleMorgenstein · 2026-11-13
- Outworld Labs launches simulation platform with new physics engine for robot learning — ycombinator · 2026-11-13
- YC-backed Mantle launches AI-native maintenance service for robots with 250+ technicians — ycombinator · 2026-11-13