K2

K² · Artificial intelligence

Cheap messy robot data becomes trainable if language critiques replace scalar scores

The authors propose a language-critique framework for imitation learning from suboptimal demonstrations that uses natural language as a structured supervision signal instead of scalar signals.

CY
DW

Chih-Han Yang, Dai-Jie Wu, Yun-Ping Huang et al.

6 authors · cs.LG, cs.AI

arXiv preprintArtificial intelligenceJul 2026 · ~55s read

The 30-second scan

Say you're teaching a robot to move a cup by showing it example moves.

  1. The method constructs language labels from demonstrations that describe current progress, identify suboptimal behaviors, and provide fine-grained corrective guidance.
  2. The authors introduce a language-critique loss that trains policies using these structured signals without reducing them to scalars.
  3. The loss is instantiated for both behavior cloning and diffusion policies, yielding two methods named LC-BC and LC-DP.
Read the originalPDF