# Information self-structuring > Information self-structuring means creating the information needed to perform a task oneself.[^1] Information self-structuring means creating the information needed to perform a task oneself.[^1] In the traditional view, the environment simply exists passively, and an agent receives information "already there" in the environment. But in examples such as and Scrabble (the word game) discussed in [The Extended Mind](https://wiki.g15e.com/en/pages/The%20extended%20mind%20(book.txt)), agents modify the environment to create the information they need for a task—or transform existing information into a form suited to the task—and then take it in. This is called . Transforming the world through epistemic action to create the information one needs is **information self-structuring**. Consider a learning robot called BABYBOT. (A summary of section 1.6 of [Andy Clark](https://wiki.g15e.com/en/pages/Andy%20Clark.txt)'s [Supersizing the mind](https://wiki.g15e.com/en/pages/Supersizing%20the%20mind.txt).) ![BABYBOT.png](https://wiki.g15e.com/data/BABYBOT.webp) This robot manipulates the world by moving its arms, learning about its own body and the world from the resulting audiovisual information. Recognizing an object placed before its eyes, for example, is harder than it might seem: it must separate the foreground object's boundaries from the background in a static image. But the problem can become easier if it can touch the object with its arm. The pattern of visual information changes at the moment the arm touches the object (from a pattern of the arm moving alone to one in which the arm and the object it touches move together). Such pattern changes make it easier to identify the object's boundaries. One important factor here is timing. The arm's movement, the resulting changes in vision, and the accompanying auditory signals (motor sounds, the sound of the arm touching an object, etc.) all occur in synchrony. This allows information from different sensory modes to be integrated (cross-modal binding) and used for learning. Bodily control must be adaptive for many reasons: changes in the body through injury or growth, extensions of the body through tool use, or constraints that arise when carrying a load. Bongard, Zykov, and Lipson 2006 reportedly created a quadruped robot that continually learned its own body image through information self-structuring. The robot could reportedly even identify its reflected body in a mirror as itself. Finally, Yu, , and Aslin 2005 show that information self-structuring can serve as a means of exchanging information between two agents. In their experiment, a person acting as a caregiver teaches words to an [artificial neural network](https://wiki.g15e.com/en/pages/Artificial%20neural%20network.txt). The network receives information produced by the caregiver through a camera, eye tracker, microphone, hand-motion tracker, and so on. With no prior model of language, the ANN reportedly learned image–word mappings solely from multiple unprocessed streams of information (visual, auditory, hand movements, etc.). The key was time-locked sensory input carrying the caregiver's embodied intention. (Related material: [The birth of a word](https://wiki.g15e.com/pages/The%20birth%20of%20a%20word.txt).) (The setting of a neural network with no built-in language model does not fit well with [evolutionary psychology](https://wiki.g15e.com/en/pages/Evolutionary%20psychology.txt). I will discuss this in a separate article later.) This last example has an interesting implication. The very strategies by which an embodied agent uses its body and the world can provide useful information to another embodied agent. In the context of [active sensing](https://wiki.g15e.com/pages/Active%20sensing.txt), for example, an agent's gaze direction conveys its current focus of interest, information another agent can use (as in the ANN example above). ## Footnotes [^1]: [Supersizing the mind](https://wiki.g15e.com/en/pages/Supersizing%20the%20mind.txt)