Robotics in-context learning sparks debate over definition of gradient updates

NielsRogge · x · 2026-08-20

Niels Rogge critiques a paper claiming success in robotic learning via few-shot gradient steps. He notes that GPT-3's in-context learning occurs without gradient updates, whereas the paper's approach of performing 10 gradient steps on 5 minutes of data implies explicit training, differing from standard definitions.

Original post →

More from Embodied

Embodied channel →