Physical-System Reality Check
For anything embodied, budget for three things — a brain, a body, and real application scenarios.
- Difficulty
- Moderate
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 82%
A framework for setting realistic timelines on physical/embodied products versus pure software. Physical systems are closer to self-driving cars than to language models: they need not just intelligence (the brain) but a working physical body and mature application scenarios, supply chains, and hardware. Use it to temper hype and forecast how long an embodied technology will really take.
Origin
Answering why 'the bitter lesson alone' (Richard Sutton's principle that simple models plus lots of data win) won't be enough for robots, Li argues robots are physical systems. She points to self-driving cars: 20 years from the 2005-06 DARPA prototype to Waymo, and 'we're not even done' — and robots are harder, operating in 3D and needing to touch things.
Core principles
- 01Software AI enjoys a clean loop; physical systems must also solve body and deployment.
- 02A robot needs a brain, a physical body, and application scenarios — all three, not just the model.
- 03Deep learning accelerates the brain but not the maturity of hardware, supply chains, and use cases.
- 04Harder physical constraints (3D, contact) mean longer timelines than any software analogy suggests.
How to run it
- 1
Classify the system
Decide whether it is pure software or a physical system. If physical, benchmark against self-driving cars, not language models.
Watch out Robots are harder than cars — 3D operation and the goal of touching things, versus 'metal boxes on 2D surfaces' whose goal is to touch nothing.
- 2
Budget the brain
Account for the intelligence/model — where deep learning genuinely accelerates progress.
- 3
Budget the body
Account for the physical hardware that must reliably act in the real world.
- 4
Budget the application scenario
Account for mature use cases, supply chains, and hardware productization — the parts a smarter model does not fix.
Pro tip The car industry's maturity is an asset embodied robotics doesn't yet have; expect a long productization tail.
In the wild
Sebastian Thrun's Stanford car won the DARPA challenge in 2005-06, driving 130 miles across the Nevada desert. Twenty years later Waymo runs in San Francisco, 'and we're not even done yet' — despite cars being simpler than general robots.
→ Sets a realistic multi-year prior for embodied AI, tempering expectations that better models alone will deliver home robots soon.
Common mistakes
Applying software timelines to hardware
Assuming robots will follow the fast scaling curve of language models ignores that they are physical systems needing body and deployment maturity, closer to self-driving cars' 20-year arc.
Is it for you?
Best for
Investors, founders, and product leaders forecasting timelines for robotics, hardware, or any embodied AI.
Not ideal for
Pure software products where there is no physical body or hardware supply chain to account for.
From the transcript
“robots are physical systems. So robots are closer to self-driving cars than a large language model.”
“we not only need brains, we also need the physical body, we also need application scenarios”
“self-driving cars are much simpler robots. They're just metal boxes running on 2D surfaces. And the goal is not to touch anything.”
From the episode
The Godmother of AI on jobs, robots, and why world models are next
Dr. Fei-Fei Li