Mark H. Lee - How to Grow a Robot: Developing Human-Friendly, Social AI

Cognition depends upon the kinds of experiences that come from having a body with particular perceptual and motor capabilities that are inseparably linked and that together form the matrix within which reasoning, memory, emotion, language, and all other aspects of mental life are embedded.

Esther Thelen (2000)

Part II of this book described a methodology and some experiments on developmental robotics. Such controlled laboratory experiments, both successful and less so, produce a range of results that are analyzed and interpreted to give insight and understanding. It may be useful for some readers, who might wish to become further involved, to bring this experience together in terms of some succinct developmental principles. Actually, not many of these are fully formed principles, but they capture and illustrate the main concepts, offer advice, and provide technical insights from hard-won research experience. Hopefully, these will be relevant for those interested in the design and building of developing robots, and even in starting their own implementation project.

The key requirements are embodiment, constructive cognitive growth, and enactive behavior.
Interactions, both physical and social, are vital for cognitive growth and shaping behavior, so bodies are essential as the substrate for perception and interaction. Mental growth requires general cognitive structures with maximum flexibility in order to support multiple functions. Action, driven by intrinsic motivation, is the force behind development.

Keep a holistic view.
Development involves the whole agent; all subsystems need to interact and work together and will be involved in all learning. This means mechanisms should be as general as possible and apply to all levels and systems.

Forget about designing robots for specific tasks.
Dont think of robot design in terms of tasks and goals. Humanlike robots will be driven by internal motivation and have a general, open-ended, and task-free learning approach. This means that designers should mainly be concerned with motivational mechanisms and general learning methods for dealing with novel situations and problems.

Employ simple modelscompact and transparent representations.
It is important that the models used are easy to analyze and relate to the results produced. This requires transparency and low complexity. Aim for the minimal and simplest set of initial principles or mechanisms and build complexity only as needed to support the current level of interaction. If you think you need to solve two problems, is there a solution to another problem that might cover both?

Play behavior is essential for the growth of skill and intelligence.
Play can be seen as the main intrinsic driver for developing robots. It can be generated through the probabilistic variation of experience applied to new situations. This is a general algorithm that covers a wide range of behaviors, from early motor babbling, through solitary physical activity, to conversing and other interactive and social activities.

Play is not complicated.
If play is seen as the modulation of prior experience to fit a current situation, then it can be implemented by remembering action schemas that match parts of the current focus of attention and trying them out with elements that are either modified or replaced by features from the current activity.

Stages are qualitatively distinct, and transitions are achieved by overcoming constraints.
Timelines and schedules give important information about when and where constraints apply. The cognitive structures will grow and change with increasing stages, but the underlying mechanism of saturation and constraint lifting is a consistent process.

The contingency rule is a powerful learning tool.
If two events occur at the same time and this contingency can be repeated, then they are very probably connected. This criterion is a major learning rule.

Start at the beginningyou cant short-circuit development.
Development builds stage by stage, so make sure that you are beginning at the most appropriate point. In other words, any preprepared behaviors or intrinsic abilities must be fully justified by careful analysis of the developmental timeline.

Start smallthen grow in resolution, refinement, and complexity.
Humans begin with very primitive perceptions and actions and then refine and combine them to create more complex structures. It is more important to get the general shape of the landscape before filling in the fine details.

Long-term memory needs to associate previous experience with current stimuli and actions.
An associative memory can operate in various ways: It can retrieve candidate actions to apply them in possible present and future contexts; it can suggest actions that might lead to desired results; and it can predict the outcome of a proposed action.

Intrinsic motivation can be implemented as novelty and curiosity.
A basic novelty detector makes for a very effective intrinsic motivator. This can be any event, action, or stimulus that is unfamiliar. Related methods for curiosity, unexpectedness, and surprise are closely related and give similar results.

Overlapping fields have many advantages over arrays of contiguous elements.
Such structures support superresolution, wherein much higher accuracy is obtained than that given by the spacing resolution of the sensing elements or pixels. They also provide compact representations of complex spatial transforms. And, most important, mappings between such arrays can be established with far fewer connections than are required with arrays of contiguous pixels. This means that both learning and using mappings can be achieved very quickly.

Be careful when using reward schemes.
Rewards can become a means of covertly setting up external goals. The play process will create goals through discovery and interaction (and scaffolding). Reward in these goal situations works well, but beware of building in reward sensitivities with a view toward accomplishing future goals.

Dont inadvertently build in any programming interfaces.
The only means of programming should be through interactive experience. Structuring the environment to select outcomes and shape available experiences (scaffolding) is likely to be the main way to train a developing robot to achieve goals.

Beware of any biological incompatibility.
Any method, however neat or efficient, that violates neuroscience or psychological knowledge is going to cause trouble later. This rules out many artificial neural network (ANN) techniques, extensive training, and all offline training. (Note that infants dont have enough skill or awake time to repeat anything hundreds of times.)

Simulation and simulated bodies are only useful for exploring ideas.
The unexpected, serendipitous, and detailed complexity of real life cannot be captured well enough in simulations to replace the real world. Any lack of authenticity in modeling the richness and interactive subtleties of actual bodily systems leads to the buildup of errors during the accumulation of experience and thought. Simulation is never 100 percent authentic, and many real-life events can only be experienced properly in the real, physical world.

Developmental robotics offers a way toward the social robot ideal.
The developmental approach, involving subjective experience and modeling of the self and others, offers an ideal test bed for building and understanding social robots. There is much ongoing research on robot/human interactions and social robotics, but this work, like AI, is fragmented across topics. Longitudinal experiments can provide the integration needed for developmental progress.