The Evolution of Robot Learning: A Path to Autonomous Knowledge Transfer

Robot learning is no longer confined to the realms of science fiction. As advancements in technology continue to push the boundaries of what is possible, robots are now being taught by one another. This intelligent transfer of knowledge has opened up new avenues for the development of autonomous systems, such as Google's self-driving cars.

From Human-Teaching to Autonomous Learning

The concept of robot learning has evolved significantly over time. Initially, robots were taught by humans through direct programming and assembly-line processes. However, the current trend is towards autonomous learning, where robots learn from their environment, experience, and each other without explicit programming. This is particularly evident in the development of self-driving vehicles.

Machine Learning: A Mechanism for Autonomous Learning

Machine learning, a subset of artificial intelligence, is the backbone of autonomous learning. Learning is defined as a mapping between an input set and an output set of symbols, where a sample maps to an acceptable response. In machine learning terms, one machine can feed pairs of values into a process on another machine, without either machine having a predefined plan or purpose for this exchange. This process is akin to teaching, but with the added ingredient of reason or planned experience, which improves the viability of the system undergoing tuition.

Examples of Autonomous Learning in Practice

Consider Google's self-driving cars. These vehicles are the result of 20 years of hard work by hundreds of engineers and scientists. Every new car that rolls off the assembly line is the result of years of assembly line automation. The robots in car assembly lines are taught to assemble vehicles, and the robots in Google's self-driving cars are taught to navigate roadways autonomously. This transfer of knowledge from one generation of robots to the next is a crucial aspect of the evolution of autonomous systems.

The Viability of Autonomous Systems

The viability of autonomous systems is determined by the ability of the system to survive and improve its capabilities. This is analogous to the concept of viability in living systems, which encompasses survival, procreation, and the ability to interact with the environment. In the context of learning, viability extends to the system's ability to adapt and improve its performance based on the environment it interacts with.

The MMARL Experiment: A Case Study in Environment-Driven Learning

At MMARL many years ago, a series of experiments were conducted to explore the feasibility of environment-driven learning. Symbolics Lisp machines served by a simulated underwater environment were used to challenge participants to build a viable underwater system that could meet specific goals. These goals were set by the programmer, and the participants were required to program autonomic responses for the underwater system.

I, however, built a system based on decision tables. This system could detect undershoot and overshoot conditions, providing feedback to the underwater system and modifying its control algorithms. This approach, akin to Hebbian learning, involved the environment providing feedback to the system, which then made adjustments based on this feedback. This process was challenging, and much of the effort was directed towards developing the decision table tools.

The key to this approach was the emphasis on high-level goals, such as survival. By focusing on goals like "don't die," the system could learn from its environment and improve its performance, thereby enhancing its viability.

Affordances of the Environment in AI

The environment plays a crucial role in the learning process of autonomous systems. In cybernetics, which is the study of communication and control in living organisms and machines, the environment provides a consistent source of sensor information. This is why AI is more about cybernetics rather than simply programming or computing. The environment acts as a teacher, providing the necessary feedback for the system to learn and adapt.

In conclusion, the evolution of robot learning has led to the development of autonomous systems that can learn from each other and the environment. This process not only enhances the viability of these systems but also opens up new possibilities for the future of artificial intelligence.