Robotics Predictions for 2025

This year was probably the most exciting year for general-purpose robotics in history. In a note that I wrote near the beginning of the year, I outlined how I expected the field to unfold, and what kinds of startups I thought would make sense. It's pretty interesting to re-read that note in the context of some pretty amazing advancements. In this post, I figured I would make a few more predictions for what I expect will happen over the course of the next year.

Hardware Convergence

Prediction: By the end of 2025, the humanoid robot will commoditized.

By this I mean, most humanoid robots will contain more or less the same components, and these components will be free for anyone to produce. Economies of scale will drive existing incumbents away from their own proprietary components and towards commodity components.

Why?

• RL-based closed-loop control, using policies trained in simulation, has firmly won as the best way of achieving real-time control.¹
• Using RL policies to control your robot means you don't need to care about a lot of properties of actuators that robotics companies have traditionally cared about or tried to differentiate themselves in.
• Most remaining hold-outs will also migrate towards some variation of the open-source MIT Cheetah actuator, once they can figure out how to make it work.
• Economies of scale amd manufacturing costs mean that alternative actuator designs will not be competitive.

Caveats

• While the hardware components will converge, there will still be some variation in the design of different robots.
  • There are still some open questions in what robot design decisions make the most sense, like how to arrange the DoFs in the hip, whether or not to incorporate DoFs in the neck or torso, and what kinds of grippers to use.
  • I suspect there will still be good reasons for different robots to make different design decisions depending on the application.

What are the implications?

• Good, general-purpose robotics hardware will become commoditized. None of the existing incumbent robotics companies, like Tesla, Figure or 1X will have any substantial advantage.
• The market price for a good, full-size humanoid robot will be below $8,000 by the end of 2025, and $4,000 for a good home robot.

No ChatGPT Moment

Prediction: There will never be a "ChatGPT moment' in robotics.

I'm defining this as some company that has been working on a model in isolation for some time, which it then releases into some commodity robotics platform, magically making that robot capable of doing general-purpose tasks with a high reliability.

Why?

• The scale of diverse data required to train a general-purpose robotics model is much larger than people currently appreciate.
• Just asking a really good language model or multimodal model to control a robot will not be sufficient. The only approach towards general-purpose embodied intelligence that can achieve "ChatGPT level" usefulness is to collect a large amount (meaning, on the order of millions of hours) of embodiment-specific data.
  • This will require shipping iterative improvements.
  • Trying to build a robot for a specific domain, then using that data to bootstrap to generality, is also a losing strategy, because of the
• The $\pi_0$ model demonstrates how far we are from having a good generalist embodied AI model.

Caveats

• Even without a "ChatGPT moment", there will be much better robots in 2025, and we will probably pass the threshold where the cost of a robot is less than the value that it provides for some large fraction of people, meaning that they will be viable products.
• We could rapidly approach ChatGPT levels of performance, but it will be incremental rather than a single breakthrough.

What are the implications?

• Some of the largest and best-funded players will need to pivot towards bringing a real product to market, or work with companies that are already doing so.

Consumer Market Growth

Prediction: By the end of 2025, more than half of the humanoid robots sold in the United States will be to consumers.

By "consumers", I mean people who are buying robots mostly for personal use, although there will be quite a lot of overlap with business uses as people experiment with them on their own.

Why?

• The utility of general-purpose robots will be quite different from the utility of previous generations of robots.
  • The "killer apps" for general-purpose robots, at first, will be speech and vision.
  • Early adopters will care more about the "humanoid" aspect and less about the "robot" aspect.
• Anyone tying themselves to traditional robotics SLAs for business customers will end up falling into the same trap that has ensnared many robotics companies in the past. They will be forced away from building good generalist models, and towards re-implementing classical control on platforms with too many degrees of freedom.

Caveats

• I think there will be business adoption in addition to consumer adoption, but most real business adoption will look more pro-sumer than enterprize, a noticable departure from traditional robotics adoption.
• Getting a new category of consumer electronics product to market is hard. Many people will waste a lot of money trying to figure out how to do it.

What are the implications?

• Robotics companies will need to start caring about branding, broad consumer appeal, consumer market segmentation, and other things that they have not traditionally had to care about.
• SLAs will become mostly unimportant. Robotics companies will start looking more like Silicon Valley tech companies, caring about things like user engagement and retention instead.