This work explores whether VLMs can achieve in-context learning (ICL) for robotics, enabling adaptation to new tasks through demonstrations and feedback without gradient updates. The goal is to translate VLM reasoning into executable robot behavior from new initial states.
HOW THIS AFFECTS YOU
●
builderYou can potentially deploy more flexible robotic agents that adapt to new environments via prompting rather than retraining.
●
researcherThis tests the limits of zero-shot generalization and ICL in embodied AI systems.