OPINION: Development of humanoid robots requires plenty of engineers to build and design the systems and algorithms.
Technology never sleeps, so it came as little surprise when media attention turned to humanoid robot startups just as the dust was settling on ChatGPT.
We had robots from Tesla, Amazon and a whole host of smaller startups. But, just like the chatbot craze before it, these robots will need huge numbers of robotics engineers and researchers to build and design the systems and algorithms that go into them.
At the start of July, I was invited to speak at the Australian School of Robotic Systems winter school conference hosted by The University of Sydney.
These are the kinds of places where researchers pass down the essential knowledge new PhD students should know before they get too far into their studies. By looking at what researchers consider fundamental today, we get a glimpse at the technology needs these robots might have in the future.
The first fundamental concept was ‘sensing and perception’, which covered approaches to how robots can understand their environment.
With the exception of radar, most of the sensors for modern robotics can be found on a high-end iPhone, such as high-definition cameras, GPS, Inertial Measurement Units and lidar (light detection and ranging).
But modern robotics isn’t just about gathering data from sensors; it’s important to know how to place the robot at a particular place, which relies on a class of algorithms called slam (Simultaneous Localisation and Mapping).
The students were then exposed to basic concepts in how to combine sensor data together, something that still requires a lot of thought, even though they may be tempted to mix everything together into a deep neural network.
Once a robot understands its environment and where it is, the second important concept is ‘reasoning and planning’. Although a human might find it easy navigating a house without walking into a wall, a robot may have trouble. This is because there are many ways a robot can be moved, and its computer needs to consider lots and lots of them to find a safe path.
For robots with arms, it gets worse, as it must consider how each joint should move as well as avoiding environmental obstacles. The more joints a robot has, the harder the problem.
This is a high-dimensionality planning problem, but fortunately lots of algorithms already exist to solve it.
Now that the robot has understood its environment and made a plan, it needs to perform the action.
This is the third important concept, ‘control and estimation’. This subject begins with the PID (Proportional, Integral and Differential) control method, which is used to control the motors that turn the joints.
If your robot has only one joint, this is easy, but it becomes much more complex when many wheels and arms are being controlled.
As the number of joints increases, it also increases the robot’s uncertainty in its body position. Uncertainty is a big concern for roboticists, but it can be confusing.
A good analogy is to consider hitting a piñata. Once the player has spun around a few times, they are very uncertain about where the piñata is, just like a robot will be as it accumulates errors over time through its joints.
The fourth topic the students learned about was ‘robot learning’. This is the use of large language models, generative AI as well as simulation and reinforcement learning to create complex behaviour.
Since robots are very expensive, understanding how to take advantage of simulation is a key skill a new roboticist will need. However, simulations are always approximations of the real world, so simulated robot agents will fail in the real world if their creators don’t take sufficient care in their simulation setup.
Most of the foundations of robotics topics covered techniques that have been available for the past ten years. Perhaps the reason we are seeing more humanoid robots now isn’t that the technology hasn’t been available, just that the software and the willingness to invest wasn’t there until the world was inspired by the rise of large language models.
• John Vial has a PhD in robotics and has spent the past several years leading teams in major Perth businesses focused on AI and robotics
