Unlock Editor’s Digest Lock for Free
For the tech industry, drunk with advances in artificial intelligence, the idea that a fully humanoid robot will soon stalke the Earth seems almost like a stretch.
Elon Musk recently forecast a $10 million market for Optimus, Tesla’s attempt at artificial humans that can take over your household chores. Nvidia boss Jensen Huang said this would become “the biggest technology industry the world has ever seen.”
And it’s easy to believe that such a revolution is at hand, to judge by the surge in investments in robot startups and a flood of online videos of two-legged robots showing impressive human-like movements. If large-scale language models can tackle difficult inference tasks, it may seem easy to port the model into a robot and retrain it to navigate the world. The problem has been resolved.
This seriously underestimates the difficulties. Thanks to decades of science fiction, many people “surprise AI is essentially embodied,” notes Playground Global venture investor Peter Barrett. In reality, bringing intelligence into the physical world is a much bigger leap.
There will be a whole new way to train a robot’s brain. When it comes to bringing people closer to powerful autonomous hardware systems, there is no room for “hatography” that LLMs tend to do today. And it doesn’t start to hurt the surface of many problems that robot manufacturers need to overcome in the construction and control of complex hardware systems designed to emulate the human body.
By raising expectations about the practicality of artificial humans, robot makers make things much more difficult for themselves than they need to. They also risk missing out on a closer, very important market that is open for robots that don’t have two legs and robots that attempt human apes in all their complexity.
In terms of artificial intelligence, robotics companies face several hurdles beyond what they face today’s LLM manufacturers. Services like ChatGPT are primarily based on models trained on the Internet, but there is no off-the-shelf corpus of data that describe the physical world.
Additionally, machines that interact with the world and manipulate objects face a much higher degree of difficulty than simple autonomous machines such as self-driving cars. Vehicles simply travel around the world without hitting anything. The robot must be able to apply touches to accomplish the most basic tasks.
There is also the issue of “planning” issues, or real-time decisions on a series of actions based on floods of real-world sensory data. This is one of the most difficult problems in robotics. Driverless cars may eventually appear on city streets, but it took years to reach this stage than the tech boosters predicted. Robots represent a much higher degree of difficulty.
At this week’s annual technical conference in Silicon Valley, Nvidia took some of these issues head on. Its COSMOS system was developed to create a virtual world that can be used to train robotic brains, but it is unclear how far this synthetic data will go when it replaces the real thing. The chipmaker also said that it has begun work on the development of a “physics engine” that will help robots understand the properties of many different things that a robot may encounter, for example, by distinguishing between stiff and soft objects. Work on the physics engine is being carried out alongside Disney and Google’s Deep Mind. This is aligning the profits of companies that speak volumes about the combination of deep technology and fantasy driving the robotics revolution.
Recommended
Nvidia has also released its early robot operating system as an open source project, which could attract other developers. It could drive the field faster, but it managed to put aside the efforts of many others who ran into the field. And laying out the development program above is far from showing actual results.
Instead of emulating people, there may be more opportunities to create boring machines that are made to handle a single task or work in use-adaptive environments such as warehouses and factories. It includes machines such as the automatic warehouse cart built by Robust.ai, the launch of Rodney Brooks startup, the founder of Roomba Vacuum Cleaner and former AI professor at the Massachusetts Institute of Technology. The dishwasher doesn’t require hands and arms to alleviate humans from boring housework. Applying the latest AI and low-cost hardware can generate useful robot waves.
richard.waters@ft.com
