Samsung's RoboICL teaches frozen GPT-6 Astra robot skills via in-context demos, no training

机器之心 · wechat · 2026-10-01

Samsung's foundation model team (lead by Ye Hui Tang) released RoboICL: instead of training task-specific VLAs, it freezes GPT-6 Astra and feeds demonstrations plus the robot's own interaction history as embodied in-context data in a shared "observation → action → execution receipt → outcome" syntax, outputting dual-arm actions directly.

Key results

Design insights

Limits: seed-0 sim runs, progress score ≠ success rate, only 3 real tasks with 5 trials each, and towel-folding tests reveal over-imitation of demo motion scale.

Original post →

More from Embodied

Embodied channel →