From a technical perspective, Artificial Intelligence research has developed algorithm for controlling robots. The most advanced one are motion planning with model predictive control. The idea is to create a forward model of the system and use the model for trajectory planning. This allows to build biped robots and manipulation robot hands.
Somebody may argue, that the practical applications are obvious. Because it's possible to utilize the technique in self-driving cars and in pick&place robots which can be used for industrial applications. There is only a smaller problem. It seems, that solving a robotics task from a technical application is not enough. That means, on the one hand it's possible to build a pick&place robot and on the other side it's not possible.
What can be solved with modern AI techniques easily are so called robotics challenges. That are synthetic challenges like micromouse, robocup, Mario AI or robot pick&place tasks. The mentioned combination of a forward model, motion planning and model predictive control results into a working system. The problem is, that a synthetic robot challenge is very different from a practical application of a robot. A practical application is equal to convert the technology into a product, and sell it to customers. Exactly this is not possible and attempts from the past into that direction have failed.
What does that mean? Creating a robot with the help of model predictive control and planning is the best practice method for building a pick&place robot which works in a challenge. Using the same technique to build a commercial robot which is sold on the market will fail. That means, the same technique is a powerful one and a useless one at the same time. This paradox is hard to grasp. Mostly it's assumed, that artificial Intelligence is a technical challenge. That means, if the algorithm was identified to control a robot than the overall problem is solved.
To measure the bottleneck in reality it's not enough to describe robotics from a technical perspective. The more elaborated form is to investigate the history of robotics companies and their failed attempt in selling a product to customers. The good news is, that in the last decade many examples are visible. The interesting question is why a certain robotics company have failed to sell the product. There are some arguments available:
- price of the product is too high. This was the case for the PR2 robot which costs around 400k US$
- robot technology is not advanced enough. This was the case for the helpmate robot in the 1990s. In that time, the onboard computer was slow and the sensors were not accurate.
- price is low and robot technology is advanced but the customer doesn't buy the product too, this was the case for the Baxter robot from rethink robotics
The pessimistic prediction is, that even the robot is sold for little money and is working with the latest hardware and software the customer won't buy the product. This outlook is equal to a general failure of robotics. Which means, that it's not possible overall to sell a robot to customers. Let us construct a hypothetical example. Suppose a company builds a pick&place robot which is very cheap and is working with model predictive control. Will this product become successful on the market or not?
The prediction is, that the robot won't find it's customers. The reason is, that even a lowcost, MPC-based robot is not able to increase the productivity in a real life condition. The task which is solved by the robot, and the requirements in reality are not the same. Or let me give another example which is in the domain of most people.
Suppose, a car company develops a self-driving car for the same price like a normal car. It's equipped with the latest sensor technology and advanced AI software. Technically spoken the car is able to drive autonomously. Will this car get customers or not? The naive prediction is, that such a car will get sold many million times worldwide. Because it reduces the workload of all the human drivers. It is useful for private households and commercial applications as well.
The problem is, that such an optimistic assumption is maybe wrong. In reality, a self-driving car is not fulfilling the real requirements. It won't reduce the workload of the driver but it increases the workload. That means, the human driver will deactivate the autopilot if he likes to relax a bit. This kind of counter-intuitive strategy isn't showing a missing knowledge of the human driver but it shows, that something with robotics technology is wrong.
Autopilots in ships
An interesting example of automated controlled vehicles is a ship. Autopilots for ships are available for decades, or at least it is written in the literature that such autopilots are available. A more realistic investigation comes to the conclusion that 0% of all ships today are using an autopilot. In contrast, all the miles were driven manual without computerized decision support. How can this mismatch be explained?
It's important to separate between autopilots in ships and the literature about autopilots. What is explained in the literature is the technical working of an autopilot. In the average book it's mentioned that steering a ship is a mathematical challenge. In some newer publication it's described as a problem for control theory which can be solved with modern algorithms. It's possible to make this problem more obvious by developing an autopilot from scratch and compare different RC controlled ships in a challenge.
On the other hand, there is the problem of ship steering in reality. In reality, the autopilot is something which isn't available. A ship is working with a human operator in the loop. The operator has some buttons he is able to press and before he can do so, he has to ask the captain what he likes to do next. An autonomous controlled ship would be equal to replace human work with software. This futuristic vision is not available in the reality, and it won't happen in the next 50 years. In reality, the amount of human operators on the bridge is constant. That means today ships are controlled the same way, like in the 1950s.
Sure, the technology has evolved a little bit. Today's ships are using computers and modern sensors. But the productivity is the same. Productivity is the measurements how many human workers are needed to control a ship with a certain size. The productivity has never increased. That means, computer technology was not able to replace human work with algorithms.
The surprising fact is, that even remote controlled ships are not available right now. It's a vision for the future to increase the efficiency for freight transport. The idea is, that if the human operator can stay outside of the ship it's easier to control the device. In case of a remote controlled ship, the amount of needed humans in the loop remains the same. The only advantage is, that they doesn't need to by physical on the ship. This kind of low end automation was never realized because the disadvantage is, that many new sensors and costly equipment has to be installed on the ship. That means, even in the year 2019, the amount of remote control freight ships is 0%. The prediction is, that in the next 50 years the amount will be the same.
Conclusion: Autopilot for ships are not available. Remote controlled ships are not available. The productivity over the last 50 years hasn't increased and it's not possible to use Autopilots in the reality. What is written in the book about Artificial Intelligence is wishful thinking.
No comments:
Post a Comment