Asimov’s Laws of Robotics and Robots Learning

Science fiction writer, Isaac Asimov, unlike most authors at the time, viewed robots positively as non-threatening, helpful robots that can be controlled by humans. [1] To eliminate the distrust of robots as evil, he along with John Campbell established the Laws of Robots which are common themes throughout his stories and have come to be widely accepted in robotics today.[2] The Laws of Robotics include; “a robot may not injure a human being, or, through inaction, allow a human being to come to harm, a robot must obey orders given to it by human beings, except where such orders would conflict with the First Law and a robot must protect its own existence as long as such protection does not conflict with the First or Second Law.”[3] These rules have later been modified.[4]

I don’t think that Asimov’s view on robots is a realistic depiction of robots today. For example, how will a robot be able to understand all the ways in which it can cause harm to a human both physically and emotionally? How can these three laws take into consideration all the unique real-world situations and how could you assure correct action by the robot?

While Asimov’s rules have become popular, they are ‘fictional devices” and should not ‘be used in real life.’ [5] The article “Principles of robotics: regulating robots in the real world” highlights the tensions and problems with these laws where there can be loopholes in the laws.’[6] Asimov’s laws suggest that robots have ‘human like’ characteristics and are responsible for not causing harm to humans. This raises the question of whether robots should be given legal personhood and what type of liability should be attributed to incidents involving robots. [7] I believe that it is humans rather than robots that should have a legal responsibility as it is humans that create, program and design robots and that holding robots as opposed to humans accountable will lead humans to take less precautionary steps in ensuring that robots are designed to behave adequately.

Many questions arise if one were to apply Asimov’s laws in our modern-day society. How will a robot be able to interpret non-verbal directions in applying Asimov’s second law? Given the high cost of producing a robot, will a robot really be designed to comply with Asimov’s third law? Given the complexities of today’s society, perhaps the best approach would be to recognize that Asimov’s laws are unrealistic and instead consider the ethical rules and values which should apply to robots.

[1] Clarke, Roger. “Asimov’s laws of robotics: Implications for information technology. 1&2.” Computer27.1 (1994): 57–66. http://www.rogerclarke.com/SOS/Asimov.html Accessed 26 September 2019

[2] Ibid

[3] Supra 1

[4] Supra 1

[5] Boden, Margaret, Joanna Bryson, Darwin Caldwell, Kerstin Dautenhahn, Lilian Edwards, Sarah Kember, Paul Newman et al. “Principles of robotics: regulating robots in the real world.” Connection Science 29, no. 2 (2017): 124–129.

[6] Ibid

[7[ Supra 5

Summary of Chapter 21 “Things Keep Getting Better Learning” of the book Robotics Primer by Maja J. Matric

Robots and Learning

“Learning, the ability to acquire new knowledge or skills and improve one’s performance, is one of the distinguishing features of intelligence, human or robotic”

Robots can learn the following:

1. A robot can learn from itself
2. A robot can learn about its environment
3. A robot can learn about other robots

Why should a robot learn?

• Through learning, robots can learn to perform better at tasks.
• Robots can adapt to changes in the environment/tasks that were not foreseen by the programmer or pre-programmed into the robot
• More simplified programming can occur as the robot can learn by itself

Learning Mechanisms

Reinforcement Learning: robots learn from feedback received based on their environment and learn what to do and not to do in different situations. This type of learning involves exploration. Robots try different actions to see what happens and determine the best combination of actions while tracking all previous actions taken. Exploitation involves using what the robot has learned previously. The balance between continuously learning by doing exploration and utilizing what is known to work is exploration vs exploitation.

Issues:

• Uncertainty: there could be errors in the way the robot senses and carries out actions resulting in incorrect learning
• Changes: environment/task changes that occur while the robot is in the process of learning may render what the robot learns invalid

Unsupervised vs Supervised Learning: In unsupervised learning, no supervisor exists to dictate what the robot does and instead the robot learns by itself. Reinforcement learning is unsupervised learning. Supervised learning involves a supervisor or teacher who at a minimum indicates to the robot what it did wrong. Neural network learning is an approach used for supervised learning and supervised learning provides more feedback to the robot compared to unsupervised learning.

Issues:

• Supervised learning can be very time consuming
• Unsupervised learning the robot “cannot compute the error from the right thing to do”

Learning by Imitation/From Demonstration: Learning is done by a supervisor or teacher and the robot repeats what it sees and does not learn from trial and error

Issues: Uncertainty exists.

For a robot to use this mechanism of learning it must be able to;

• focus on the demonstration
• determine what is relevant vs irrelevant
• “match the observed behaviour to its own behaviours”
• make adjustments from what is observed
• recognize the goals of the demonstration

Learning and Forgetting

• Purposeful forgetting is important in machine learning — robots can discard outdated learned information that is no longer correct and apply more recent useful information.
• Other learning mechanisms; memory-based learning, evolutionary learning, case-based learning, statistical learning and more.

Thought question: How can we overcome the uncertainties faced by robot learning?

Perhaps introducing a fuzzy logic-based strategy could be a means of overcoming uncertainties in robotic learning. Fuzzy logic can help the learning process, decreasing the number of random actions taken by robots.[1]

[1] Ierache, Jorge “Application of Fuzzy Logic in Learning Autonomous Robots Systems” (2013) Advances in Intelligent Systems and Computing Vol 208