“That’s it, it’s working! », I sent to my entire repertoire, with supporting evidence. After months on all fours, weeks in a “crab”, my baby finally took the plunge at 13 (and a half) months. He walked. For any parent, this non-achievement is a milestone in the baby’s motor learning. The fear – a little irrational – that one’s offspring will remain stuck in an intermediate position, standing, clinging to furniture or a walker to move forward, can finally disappear in favor of new anxieties.
While the new generation of artificial intelligence such as ChatGPT has given our toddlers a hard time in the area of language learning (we’re matching here), where do humanoids stand in the walking field? Are Boston Dynamics’ famous robots more competent bipeds than humans or, on the contrary, is humanity finishing in the lead?
Let’s start by analyzing how the baby goes from a limp body at birth to a little person capable of standing up and running a few months later. “Each stage sets up the connection networks which help it move towards the next stage,” explains Séverine Alonso-Bekier, psychomotor therapist. The newborn is very soft. He first holds his head at around three months before being able to hold his back, then sit up between 6 and 8-9 months.” He gradually strengthens his pelvis by lifting his buttocks in the four-legged position before toning his little legs.
The robotic appearance of the inverted pendulum
Casually, putting yourself in a position does not require the same skills as simply holding a position. Going from flat on all fours to sliding into the seated position uses other muscles than just holding the seated position. Although my baby got up early (at 8 months), he sat up on his own quite late. And let’s not talk about the four-legged position in which he deigned to take an interest at just 10 months old, preferring to crawl military style in the trenches for almost a semester. Overall, let’s be reassured, all babies end up walking, whether or not they skip some of these steps.
On the machine side, learning is very different. Already, unlike a baby, we do not release a humanoid robot worth thousands into the wild to let it learn on its own with the risk of breaking the prototype. The machine learns on a computer, in a simulated universe where its rigid frame poses no risk. “The inverted pendulum is the simplest model that can be applied to the robot,” explains Ewen Dantec, postdoctoral researcher in robotics at Inria. A mass at the end of a stick falls with a certain consistency of time and it is enough to add a second stick at the right time – which symbolizes the second leg – to catch the mass. It’s not a very big mistake to say that humans move like an inverse pendulum.” And this method works quite well, even if the appearance is not very human.
With the increase in computing capacities, research in reinforced learning develops algorithms that allow the machine to learn from its errors. Robots learn on their own, a bit like our children who experiment, bump into things and fall before becoming little walking pros. “We start from zero, the machine knows nothing. We make her do a lot of trial and error so that she can stand up and walk,” explains the scientist. “It’s very recent. We take a robot, we make it simulate its behavior 2000 times with random actions and we try to improve its behavior based on the reward it receives. We see parallels with babies in the sense that when they start to move, it’s very risky, they try things, and as they are very soft, they don’t hurt themselves. After a while, by trying a series of precise movements, they will succeed in moving forward.”
A battle lost in advance?
That’s all well and good, but the transition from virtual to real is not always easy. When the robot tries the same thing IRL, sometimes it is no longer capable of anything. A point for the baby. “That’s what held us back for years. The simulations worked perfectly, and, with the robot, nothing worked anymore. We solved this problem by introducing randomness into the simulation, continues Ewen Dantec. In the environment that the robot perceives, we add a noise, as if it was seeing blurry or as if it could not feel its joints well. This increases its robustness.” It’s a bit like humans, if they see blurry, they will be more vigilant. He slowly moves his hand towards an object and once reassured, he continues on his way. “The robot will be more careful in its movements and when we put it on the real platform, it will work better,” he points out.
On the latest graphics processors (GPU), “we can train a simple walking policy in a few hours,” admits the robotics researcher. A hard blow for the baby who will have difficulty putting one foot in front of the other before several months of learning. On average, fifteen months, according to my son’s pediatrician. And again, we’re being nice, because it takes three more months to consider confident walking that resembles that of an adult. On average, eighteen months for a baby, compared to a few hours for the machine. Hard blow to the ego of humanity. But the fight is not fair.
“We have monstrous machines in terms of computing capacity. It’s as if the baby had a thousand brains on which it could test lots of different walking policies, puts Ewen Dantec into perspective. Obviously, it goes faster.” Of course, the machine learns in a simulated universe. Certainly, the models still lack robustness. But we’re not going to lie, the robot’s learning speed frankly puts my enthusiasm in front of my baby’s exploits into perspective.