AI Robotics – Moravec is Alive!

The domestic house bot is still a pipe dream. 

• I do not doubt that there is a very large market opportunity to sell robots that can do the housework and many other things but the problem is still so large, that no one is going to get there anytime soon regardless of the hubris and hype being pushed by the AI companies.
• Google is following Meta, OpenAI and Tesla in chasing the robotics opportunity but while the proponents talk a good game and produce amazing videos, reality is still a long way behind.
• Google has launched two models Gemini Robotics which can take language or image inputs and then output actions and Gemini Robotics-ER which enables a robot to perceive and understand the real 3D world.
• Google demonstrated a series of impressive tasks like putting small objects in dishes that are being moved around and putting a slice of bread into a ziplock bag.
• However, what Google and everyone else keeps fairly quiet about is that all of these tasks have to be specifically taught to the robot and the algorithms that run the motors that generate movement are unique to every single robot.
• This brings us right back to the usual place which is causality.
• All of these models are based on a system that uses statistical pattern recognition as the method for making decisions, and as such, they have no real understanding of the task they are executing.
• A good example of this problem is the very high likelihood that if the colour of the Ziplock bag is changed from transparent to red, the robot will fail to complete the task.
• Even the dimmest human would not fail to make that adjustment.
• It is not until the robots can make these adjustments in real-time and truly understand the causality of what they are being asked to do, that robots will become a huge mass market.
• Google describes the criteria for a helpful robot as interactive (i.e. you can talk to it and it can adjust for an object’s movement), dextrous (solvable with engineering) and general which is where the whole proposition falls over (see above).
• Google also states that all of these capabilities need to work across different robot form factors which today remains impossible.
• Google is claiming that Gemini Robotics is bringing all of these together implying that it knows how to solve these issues which I suspect is very far from the case.
• These problems are neatly summed up by Moravec’s Paradox which describes this as things that are easy for humans and animals are very hard for machines and things that are easy for machines are very hard for humans and animals.
• This is on display everywhere where the most advanced LLMs can now crush PhD level maths but fail to play tick-tac-toe or draw a picture of a human writing with their left hand.
• In robotics, this problem is even more intense because the limitation exists in both the cognitive systems of the robot and in the algorithms that control its behaviour in the real world.
• This means that while it is possible to build a robot that can walk on legs that can fall over and get back up again, the minute the system is changed or upgraded, the whole thing needs to be trained again from scratch.
• In Google’s defence, the basketball demonstration is quite impressive.
• This is where the robot picks up a small ball and puts it through a miniature basketball hoop that it has never seen before.
• The Gemini model has been trained on the concepts of basketball meaning that it has some knowledge of the task and so I am going to give Google some points for this demo.
• The dream for this sector (and I suspect of many householders who can’t afford a full-time human housekeeper) is to build a humanoid robot that can do all the housework.
• This could be a big market as I suspect households would spend as much on a household robot as they do on a vehicle.
• However, there are two big problems as LLMs are not nearly close to being reliable enough and there is still no way to train one algorithm and have it able to control many different form factors.
• Both of these need substantial progress before the robots are sufficiently reliable and cost-effective for mass manufacturing. Moravec’s Paradox is alive and well today as it was when it was first described 40 years ago.
• I think that there is still a long wait ahead, but the rewards will be big.