A humanoid robot can barely lift more than its own arms. What does it have to do with heavy machinery – and why is an industrial company like Konecranes training one?
At first glance, a robot with a human-like body makes absolutely no sense in industrial setting, says Markku Häivälä, R&I Director at Konecranes.
“It would be better if it were, for example, four-legged or had wheels, so it wouldn’t fall over easily.”
But the walking humanoid robot has something else: the best “brains” on the market. The robot, Unitree G1, has advanced neural network running capabilities and software stack, powered by Nvidia’s AI computing.
It is its ‘brain’ that counts.
“Any engineer knows that it’s bloody difficult to program a walking robot,” Häivälä says.
The humanoid form as such doesn’t matter.
“It is its ‘brain’ that counts. It’s a research platform for technologies such as reinforcement learning and programming, VR, AR, and digital twins.”
Konecranes is among the industrial companies at the forefront of testing and training humanoid robots.
Automotive companies have been running pilot tests with robot companies, but despite all the media attention, humanoid robots are still in their infancy – real commercial use cases remain limited.
How to train a robot?
In an industrial testing hall research engineer Timo Lundstedt puts on VR glasses. He waves his hand. The robot waves simultaneously, almost with no delay.
This is one way for the humanoid robot to learn: mimicking human actions. This is called teleoperation. Based on whatever the robot does after Konecranes’ research engineers, it can program itself.
In the future, robots might be able to watch video tutorials for learning new tasks, in a similar way humans might watch YouTube tutorials for changing car tires or making heatless curls.
Another way for the robot to learn is what Häivälä calls reinforcement learning: it gets positive feedback each time it does something right. Training a robot is similar to training a dog.
For each movement it learns, the robot is provided with a pre-trained neural network – a type of AI model. Research engineers write the code to combine these individual skills into coordinated motion sequences, creating an orchestrated system.
This way the robot is able to learn entire flows of movement or multi-step tasks. Things that are natural to humans, like waving a hand, take a lot of effort. Training the robot to be able to wave its hand took 3 months from Lundstedt and his colleagues.
“The robot can be good at a certain task, but humans will stay unmatched in their skills for a very long time,” Lundstedt says.
Lundstedt, Iina Lumme, and Sami Terho work side by side training the robot. The idea is for each to learn, explore and be able to put the combined know-how into use in different applications.
Any skill set one robot attains can be later copied to other robots. But a waving team of humanoid robots is not what Konecranes’ experts have in mind.
Making it matter
Working with autonomous robots is a bit like a moonshot.
“Working with autonomous robots is a bit like a moonshot,” says Markku Häivälä.
“Going to the moon is still not commercially viable, but developing the technology led to countless everyday innovations, from vacuum cleaners to memory foam.”
What interests Häivälä and his team is how to use this way of learning to develop applications for Konecranes’ existing equipment and services.
According to Markku Häivälä, the gap between all the hype around humanoid robots, and what is actually plausible and applicable, is wide.
The end goal might as well be automation, autonomous robots or the unmanned, so-called dark factory. But what’s more appealing is what’s already here or just around the corner, like AI-enhanced programming and testing environments for software, or developing digital twins to improve the efficiency of industrial production and logistics.
Of course, even in many industrial facilities there are circumstances where a humanoid form could come in handy: hazardous environments such as nuclear plants or foundries, for example.
Or, the robots could work side by side with humans and heavy machines. When there’s a 20-ton load in the air, safety is priority for Konecranes and its customers.
“A robot could walk or roll ahead of the load and inform the crane if it spots a person or a vehicle in its path. A very simple solution, but it could save a life – and also keep the customer’s business up and running,” says Häivälä.
In the longer run, the crane itself might be able to survey its surroundings and act safely based on that.
“Our machinery having some of that autonomous decision-making skill – that would definitely make me go wow,” says Häivälä.
Best part of it
Looking at what the team currently is working on, this scenario might not be science fiction at all.
“What interests me the most is how this robot can perceive and understand its surroundings by using sensors, a lidar laser scanner and a camera,” says Senior Specialist Sami Terho, who works at the intersection of computer vision, deep learning and industrial automation.
Terho drives research collaborations with the University of Helsinki and Aalto University to apply computer vision and AI. His colleague Timo Lundstedt has studied possibilities of automated smart cranes in a factory ecosystem.
At the end of the day, it’s just a machine. What’s interesting is the learning curve of the expert team working with it.
That said, still a lot has to happen before Häivälä gets his “wow!” moment. There are no manuals for humanoid robots, and hardly any benchmark data.
“For me, the best part of the work is the teleoperation, teaching the robot with the help of VR glasses. The toughest is that there is no ready documentation for anything,” says Iina Lumme.
Why has Konecranes chosen to do the heavy lifting and lead the way as a pioneer?
As a technology leader, Konecranes has the culture and appetite for innovation, says Häivälä.
“Our people are always hungry for any new stuff.”
“At the end of the day, it’s just a machine. What’s interesting is the learning curve of the expert team working with it.”
