Embodied AI Turns to In-Context Learning as Multimodal Context Becomes a New Scaling Dimension

The competition over long-context reasoning is making its way into physical systems. Early-stage teams are exploring multimodal In-Context Learning (ICL) as an alternative scaling dimension for embodied robotics, feeding models sustained streams of sensory and action history so they can adapt to unencountered tasks at runtime. Beyond expanding parameter counts and raw offline datasets, processing continuous temporal context is emerging as a pragmatic lever for robotic adaptability.