SJTU's GSR Method Fixes Robot Failure on Instruction Paraphrasing

jiqizhixin · x · 2026-08-16

Why do robots fail when instructions are rephrased? Researchers from Shanghai Jiao Tong University and Anyverse Dynamics present Grounded Semantic Re-binding (GSR) to solve this.

The issue is architectural: current models fuse visual and text data too early, creating noise. GSR extracts the task's meaning separately and fuses it with visual data only at the end, training a new action decoder from scratch. On the LIBERO-Para benchmark, GSR boosts success rates by up to 44.6%. Lightweight models now rival massive pretrained systems, setting a new PRIDE score record of 70.4. They also launched ParaVLA, a tiny 0.33B model.

Original post →

More from Embodied

Embodied channel →