Physical Prompting

Simple Definition

Physical prompting is showing a robot a physical demonstration that acts as context for performing a task.

This is an emerging term. It describes a real and increasingly common practice in robotics, but it isn’t yet standardized industry-wide the way older machine learning terms are.

Plain-English Explanation

Instead of programming a robot’s movements step by step, a person can simply demonstrate the action: moving an object, opening a container, arranging tools, or performing some other short manipulation task. The robot then attempts to reproduce or adapt that behavior based on what it just observed. It’s closer to showing someone how to do something than writing them an instruction manual.

Important Distinction

Physical prompting does not necessarily change the model’s weights. A demonstration can simply remain inside the robot’s current context, shaping what it attempts right now, the same way a text example shapes what a language model outputs right now. That makes it a physical, embodied version of in-context adaptation rather than something deeper. If a system later updates its weights based on that experience, so the lesson persists beyond the current session, that moves it closer to continual learning or fine-tuning, a meaningfully different kind of change.

Example

A warehouse robot is shown, once, how to pick up an irregularly shaped part and place it in a bin at a specific angle. Instead of an engineer writing precise coordinates and movement instructions, the robot uses the demonstration itself as its guide for the current task.

Why It Matters

Physical prompting lowers the barrier to directing robots for people who aren’t robotics engineers. A demonstration is faster to give and easier to adjust than a hand-written program, which matters as robots move into more varied, less predictable environments where hard-coding every action in advance isn’t practical.

Related Terms

• In-Context Adaptation, the broader concept this is a physical, embodied instance of
• Continual Learning, what physical prompting becomes if the demonstration actually updates the model’s weights
• Multimodal AI, the broader category of AI systems that work across more than just text
• Computer Vision, the perception capability a robot needs to interpret a demonstration in the first place