“And then we get the robots”
Progress in robotics isn't just an intelligence problem.
Most of the narrative accounts of the intelligence explosion predict very fast robotics progress will follow the invention of powerful AI. (See here, here, and here.) But none of these predictions are specific about exactly what needs to be solved. They gesture at data bottlenecks but ultimately abstract away the challenges to robotic progress and imagine that superintelligent AI could “solve robotics”. If it were possible to do this, it could well lead to much more dramatic labour automation and explosive economic growth. So we need a better understanding of how sensitive robotics progress is to AI progress.
There are a few Silicon Valley companies aiming to build humanoid robots at present. Their robots are made from electromechanical actuators, gearboxes, and cameras (though some also have LiDAR or proprioceptive sensors embedded). While these robots can produce impressive demonstrations, they cannot behave reliably in general environments. There are not a large or diverse enough set of training data for them to learn general behaviour. Language models have the luxury of an enormous dataset to make into a prediction task, where robots do not. Reinforcement learning is much cheaper than text than robots. “Generating a trajectory” means trying to solve a maths problem on a digital trackpad, not trying to drive a car. Failing is cheaper too. Language model hallucinations are quaint, self-driving car hallucinations…aren’t.
There are three ways to solve the data problem. The first is simply to gather more data from robots trying to solve problems. Google tried to solve this by building an “arm farm” and having these robots attempt to solve problems around the clock, but this was closed. Others have had humans complete tasks using robot grippers or had humans teleoperate robots to gather data. Aside from increasing the amount of data, researchers can improve the training procedure, so the robot “learns more” per example. The chart below shows how, over time, equivalent performance on an image recognition task required less data and computational resources.
Otherwise, it could become possible in the future to create simulated environments for the robots to train in. There are academic examples of researchers teaching robots to perform very simple tasks using synthetic data, but it is difficult to create realistic environments for robots to learn more complicated tasks. Google DeepMind and OpenAI are trying to create physical world models. DeepMind’s model, Genie 2, was trained on video game data to turn an image into a consistent 3-d world for up to a minute. There’s a large amount of compute scaling and data which could improve these models, but of course, it's uncertain how far this would go: it is much easier to imagine these very large models helping with planning tasks and coarse movements than fine motor control.
How could automated AI progress help with the data problem?
Big jumps in AI capabilities could accelerate the solution to the data problem in a couple of ways. Most directly, automated researchers could design better training algorithms, new model architectures and so on. However, we cannot assume that the rate of algorithmic improvement in AI researchers can generalise to robots. Researchers have to test whether their changes have worked, and testing whether an automated AI researcher is better can be done entirely within a computer whereas knowing if the automated robotics researcher can design better algorithms requires real-world (expensive) feedback. This could, however, be slightly mitigated. Automated researchers could help to design better simulated environments for the robots to train in. All of this automated software progress is subject to the same constraints as automated AI research.
But even if superintelligent AI could generate entirely realistic simulations and find the optimal learning algorithm or model architecture for the current set of hardware, robots wouldn’t be able to complete all the tasks humans can. We would run into hard limits. From a recent Construction Physics article on robot dexterity:
Human hands are very strong while being capable of complex and precise motions, and it’s difficult to match this with a robot hand. Robot hands are often surprisingly weak. An average man has enough grip strength to lift 40 kg or more off the ground (20 kg in each hand), and a strong man can lift upwards of 100 kg. By contrast, NASA’s Robonaut 2 hand had a payload capacity of 9 kilograms, and the Shadow dexterous hand (billed as the “most advanced 5-fingered robotic hand in the world”) has a payload capacity of just 4 kilograms.
More importantly, human hands are extremely sensitive, and capable of providing a lot of tactile feedback to help guide our actions. A human hand has around 17,000 touch receptors, and is sensitive enough to discriminate between textures that differ by mere nanometers. Robot hands are getting better, but still don’t appear to be close to what a human hand can do. This robot hand, for instance, boasts “17 tactile sensors,” and this one from Unitree has 94.
Optimising the current hardware would unlock some economically-useful tasks but not the “100% of human tasks” that predictions of the intelligence explosion would require. Getting closer to all physical labour being automated would require a leap forward in hardware progress.
How sensitive is robotic hardware R&D to intelligence?
ARIA has a research programme dedicated to improving robotic hardware which can provide a guide to the kinds of step-changing inventions necessary.
The next step for robotic sensing is to improve tactile sensing. One way to think about sensing is that it has been free riding on improvements in cameras and so vision capabilities far outpace other modalities. The ARIA programme has funded three tactile sensing projects. One group has developed a new material that conducts electricity in proportion to the force being applied to it, and will combine this with directional strain and temperature sensors into a single e-skin. Another group has developed a material which is continuously conducting and uses changes to the voltage to identify force on the surface. A final group is using changes in magnetic fields to allow for very granular sensing.
Electromechanical actuators are limited in a number of ways. They have low torque density, limited bandwidth (reaction time), high inertia, and they scale down very poorly. In the limit, a DC motor becomes a heat machine. The ARIA programme is funding alternative muscles:
For very fine control (e.g. finger joints or semiconductor assembly) one group has developed a material layered with liquid droplets and electrodes between the layers. When an electric current is applied to the electrodes, it causes the liquid droplets to compress, creating a “gripping” effect when the material is stacked in layers.
A group is developing a braided material for pneumatic (air or liquid) muscles that can channel the radial force during “contraction” to make control more precise.
A group is making a new material geometry for an elastomer. The material contracts like a muscle when current runs through it but this currently requires very high voltage and so the project is aiming to bring this down.
A related project is developing a muscle mimic which moves fluid around in a soft pouch.
Finally there is a group working on synthetic muscle fibres.
There are three projects which aim to reduce the number of gearboxes a robot would need:
One project is replacing the gearbox with an arrangement of magnets at different polarities to control rotary motion.
Another project is miniaturising an existing actuator that has pairs of magnets controlling linear motion.
A third project is developing clutches which could route power from a single gearbox to reduce the need for every joint to have a gearbox.
From the outside, very powerful AI would be very useful for aspects of the R&D process but wouldn’t “solve” the problem end-to-end. Many of the processes require materials R&D and AI is very useful for discovery and modelling behaviour. But these processes ultimately depend on real-world experimental data to train the models and to refine the search. Similarly, for the sensing projects, very capable AI could suggest topologies for spacing the sensors, to account for wiring complexity, the flexibility of the material, cost, and so on; but computer simulations would be an imperfect substitute for seeing how robust the material was to a month’s intense use. Very good AI should minimise, but not totally eliminate, iteration cycles. Prototyping and manufacturing for real-world experimentation becomes binding.
Were robotics progress going to happen very quickly, all of the tasks involved in hardware R&D would need to become “intelligence problems”. The crux here is to what degree do humans need to be involved in the iteration cycle: how good can the physics simulations get, that hardware design happens entirely in silico? Can the scientific agents figure out, say, the optimal arrangement of magnets and their strengths to control a robotic “shoulder” while also trading off weight, manufacturability, durability and so on? Are all of the questions answered by simulation?
The idea that a technological singularity will occur after we automate AI research abstracts away these practical bottlenecks in the R&D process. AI is going to change everything, but it won’t be overnight.
If we assume intelligence explosion, all robotics becomes no harder than teleoperated robotics. Like, humans are pretty good at using an arbitrary teleoperated robot, after hours not years of practice. If AIs are not substantially better at using an arbitrary teleoperated robot than humans are, then … that was not a true intelligence explosion! Right?
But you seem to have a different mental model, where the intelligence explosion creates AIs, which in turn do robot algorithm R&D, instead of them just piloting the robots directly … or something like that?
Anyway, I’m not an expert on teleoperated robot hardware, but I tried to look into it briefly once, and got the impression that they are much more capable, affordable, and easy to mass-manufacture than you’d think from studying the non-teleoperated robot industry—despite extremely low volumes right now, for obvious reasons. For example, I once saw a video of a teleoperated robot cleaning a messy house (folding sheets, clearing the table, etc.), I don’t remember the details but I’m pretty sure it was from 10 years ago or more, maybe much more.
(In this comment, I am severely downplaying how crazy an intelligence explosion would be, for the sake of argument.)
Good post! I made a related (much less detailed) argument a while back that physics will still exist and will presumably pose SOME kind of limit.
"""A super-intelligence wouldn’t be a god. I would expect a super-intelligence to be better than humans at creating better super-intelligences. But physics still exists! To do most things, you need to move molecules around. And humans would still be needed to do that, at least at first (https://dynomight.net/smart/)
So here’s one plausible future:
1. Super-intelligent AI is invented.
2. At first, existing robots cannot replace humans for most tasks. It doesn’t matter how brilliantly it’s programmed. There simply aren’t enough robots and the hardware isn’t good enough.
3. In order to make better robots, lots of research is needed. Humans are needed to move molecules around to build factories and to do that research.
4. So there’s a feedback loop between more/better research, robotics, energy, factories, and hardware to run the AI on.
5. Gradually that loop goes faster and faster.
6. Until one day the loop can continue without the need for humans.
That’s still rather terrifying. But it seems likely that there’s a substantial delay between step 1 and step 6. Factories and power plants take years to build (for humans). So maybe the best mental initial model is as a “multiplier on economic growth” like all the economists have been insisting all along."""
I got a fair amount of pushback along the lines that AI might exploit some unknown vector and leap to powerful robotics quickly. I guess you can't totally exclude unknown unknowns, but I still think the scenarios you lay out are most likely.