One of the 'side-effects' of our approach was that the VLM-orchestrator could be used for one-shot, in-context learning to bootstrap RL post-training with a feasible solution. The latest releases of @SkildAI's S1 and @GeneralistAI's GEN 1.5 show how powerful ICL gets at scale.🚀
Replying to @sukhijabhavy
When we started this work one year ago, there were no in-context-capable VLAs, yet. Instead VLM-orchestration for semantic exploration was our approach to go the 0->1 step, before RL can take the policy to 100. [3/4]
