Training robots to perform reasonably simple tasks requires a huge engineering effort.
Once found only in the realm of science fiction, believable AI chatbots are now installed on every Windows computer and are becoming embedded on websites across the internet.
At the same time, we’ve had a resurgence in humanoid robot companies promising robots that can perform household chores.
This is not a new promise.
In 2011, I was presenting my research at a robotics conference in San Francisco, during which we were treated to a lab tour of a company called Willow Garage.
Their founder confidently told us that his robot, the PR2, a bulky, wheeled unit with two large but compliant arms, would pave the way for a household robotics revolution.
Three years later, Willow Garage closed its doors, having been unable to find a market for its robots.
Although the company folded, its software was open sourced and is still used by robotics companies today.
So, what has changed in the 10 years since Willow Garage shut down?
Are chatbots like ChatGPT the missing piece that will finally enable the widespread adoption of robotics?
And can we use these robots to replace the jobs that humans don’t want to do?
For much of the past decade, robots were restricted to highly choreographed environments.
Roboticists would handcraft and finetune their environments and sensors for a specific setting.
Situations where that wasn’t possible, such as self-driving cars, required huge amounts of engineering effort to find all the unexpected situations and make sure the robot could operate safely.
Chatbots’ main contribution (and their underlying large language models) to robotics is in their combination of natural text with visual understanding.
It’s now possible to create a 3D scan of your house and then find the closest point in space to the text request ‘get me a beer’.
This automatic annotation of 3D context significantly reduces the human effort required to set up a robot in a new location.
Unlike chatbots, robots don’t have large data sets that can be used to train big neural networks.
So, although a robot can now know what it needs to do in a new environment, actually performing that task is still difficult.
There have been some advances in training robots during the past decade, including the development of a field called ‘reinforcement learning’.
The most successful of those techniques involves mimicking a human repetitively performing a particular task.
This means humans still need to do the task, sometimes hundreds of times, to effectively train these robots.
Demonstrations of humanoid robot by companies such as Tesla or Figure AI use these mimicking techniques to get the motions.
In reality, though, how close are we to these robots replacing manual labour?
For warehouse workers, humanoid robots such as Digit from Agility Robotics look like strong contenders.
These legged robots are able to carry boxes weighing 15 kilograms.
Others, including robots like Atlas from Boston Dynamics, have been seen grasping heavy, awkwardly shaped automotive components with their humanoid hands and stacking them on shelves.
In the kitchen, robots are starting to achieve simple cooking tasks, such as flipping an egg or cleaning up after spills.
However, wheeled robots have been part of the automation of warehouses for a long time, and egg flipping or cleaning up after spills aren’t particularly valuable tasks in this context.
Still, the tasks in these industrial settings may be easy for a human but remain challenging for a robot.
The thousands of wheel valves across industrial sites pose a difficult challenge for robots, and the simple task of fitting a large hose to move liquid between two tanks is a difficult perception and coordination task that humanoids robots will struggle with.
And while there are opportunities in resources or agriculture, given the many dangerous and repetitive tasks in these sectors that could be automated, it seems our industrial jobs are safe from the humanoid robot revolution. For the moment at least.
- 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
