INSTANT POLICY: IN-CONTEXT IMITATION LEARNING VIA GRAPH DIFFUSION

Technical Analysis learns new tasks instantly (without further training) from just one or two demonstrations, achieving in-context imitation Learning (ICIL). It introduces inductive biases through a graph representation and model ICIL as a graph generation problem with a learned diffusion process, enabling structured reasoning over demonstrations, observations, and actions. High-level Analogy: Imagine you want to teach a robot a new kitchen chore, like 'open the fridge door.' Instead of showing it hundreds of times for every type of fridge and telling it exactly how to move each joint, you just show it one or two examples of opening any fridge door. The robot then doesn't just copy the exact moves; it understands the idea of 'opening a fridge door' by learning how to connect what it sees now (its surroundings and the fridge) with what it saw in your examples (how the door moved). It's like teaching a child the concept of 'door opening' rather than just memorizing a specific sequence of muscle movements for one particular door. This understanding comes from building a mental map (a 'graph') of the task, which it then refines like an artist sketching an idea (a 'diffusion process') to figure out its next move, even for a fridge it's never seen before. Motivation of the Work The primary motivation for this work is to bring the efficiency and adaptability of In-Context Learning (ICL) from large transformer models to the field of robotics, creating what they call In-Context…