VLM orchestrator enables in-context learning to bootstrap RL post-training in robots

oanacamb · x · 2026-09-02

A robotics researcher notes a 'side-effect' of their approach: the VLM-orchestrator can do one-shot, in-context learning to bootstrap RL post-training with a feasible solution. When the work started a year ago, no ICL-capable VLAs existed, so VLM-orchestration for semantic exploration handled the 0->1 step before RL polished the policy. Recent releases — SkildAI's S1 and Generalist AI's GEN 1.5 — show how powerful ICL gets at scale.

Original post →

More from Embodied

Embodied channel →