Some current thoughts on embodied AI, written down so that I can be proven wrong later.

A note on scope: by embodied AI I mean technology meant to replace physical labor. Humanoid robots built for show are not what I’m talking about here.

• Embodied AI is fundamentally driven by data plus algorithms. The ultimate goal is to replace all physical labor, so the data has to cover the operational common sense of every industry, and it has to express the corresponding action information at high quality. The algorithm determines learning efficiency — only an algorithm that can fully digest all that data (i.e. one that scales properly) will let the system actually handle the full range of application demands.

• Data

  • Data is the single most decisive factor in whether embodied AI succeeds. There are really only two questions: what counts as good data, and how do you get it at scale. Data quality depends on coordination across hardware, operations, infra, and algorithms — no single team gets to decide it alone.

  • In Will Embodied AI Arrive Before Autonomous Driving? I wrote that embodied AI has no “robot internet” to crawl. I now think ego-centric data is what plays the role of that “internet data” — it is embodied AI’s common crawl. It’s dirty, and it can’t be tied directly to any particular robot body, but it covers a huge range of tasks and edge cases, and it supplies the basic information about scenes and tasks. Ego-centric data also needs cleaning and labeling before it delivers value, and the details of how you do that are the core of the technology.

  • UMI is embodied AI’s fine-tuning data: its input format is kept as close as possible to what the robot sees at deployment time. More importantly, it’s the only kind of data whose marginal collection cost trends toward zero as deployment scales — the person who most wants the robot to do the job well is exactly the person holding the collection device.

  • Because embodied data does circulate between players (the upstream and downstream industries are fairly hot), if any company makes significant progress with UMI, the data industry and academic research will both converge on that specific form factor — indirectly setting the standard for what embodied robots look like.

• Algorithms

  • Embodied algorithms have many parts. The action model gets most of the attention, but the agent layer and the motion control layer matter just as much.

  • The action model is a large model carrying action information, and it needs to train on many different data types together — which is a natural advantage for the big model companies. But what embodied AI delivers is a complete integrated system: hardware, motion control, several models working in concert, and behind all of it the barrier of data collection. Those are things the big model companies haven’t accumulated yet.

  • A MoE model with 200B total parameters is enough to support a genuinely usable action model. Most tasks don’t require deep coding or reasoning knowledge — it’s mostly physical common sense — and there’s a hard constraint on inference efficiency. The big model companies’ moat at trillion-parameter scale doesn’t buy them much of an advantage here.

• Where things stand

  • Because embodied AI still depends on the “body,” its adoption is inherently tied to selling bodies. In industrial deployments, embodied AI is a sales-heavy business: early models can’t work out of the box across arbitrary tasks, so right now everyone has a shot and the market is fairly fragmented. But the “embodied AI” at this stage is really just a robot arm that’s easier to “program” — not general intelligence.

  • What embodied AI is after is replacing most physical labor — or, put another way, having a GPT-3 moment for manipulation tasks: most tasks work out of the box, and the ones that don’t can be made to work cheaply through something like prompt engineering. Once that level is reached, ordinary consumers can simply buy one and use it, and that’s the starting point of real adoption. Who ultimately wins comes down to who closes the positive loop on the consumer side first: brand awareness → consumers buying → more data → a better product experience → stronger brand awareness. From some companies’ PR we can already catch a glimpse of general-purpose embodied AI, and given the pace of progress in LLMs, there’s room to be optimistic.

  • What is certain is that doing a lot of industrial applications doesn’t generalize to arbitrary tasks. Industrial scenarios are fairly narrow, and they don’t feed back any data from outside the industrial setting — which is why I stay cautious about much of today’s industrial deployment. There may be application scenarios diverse enough, and close enough to ordinary consumers, to generalize to arbitrary tasks — the hotel scenario MORPHI Robot has mentioned may well be one of them.

  • The end state is probably a large number of embodied AI companies serving industrial applications, plus a small number of home robot companies serving ordinary consumers.