TRAIL at CVPR 2026
3. Juni 2026, von Theresa Pekarek Rosin

Foto: UHH Knowledge Technology

TRAIL was represented by Theodor's findings poster at this year's CVPR in Denver, Colorado.
The paper "Grounding Hierarchical Vision-Language-Action Models Through Explicit Language-Action Alignment" addresses a persistent challenge in robot learning: how to align the chain-of-thought reasoning of a hierarchical VLA with physical actions. The proposed method uses preference learning to create explicit semantic links between language steps and robotic behaviour.
The work sparked lively discussions with researchers from around the world.
Read it here!
Project page: https://theodorwu.github.io/GPLA/

