Physical AI at CES 2026: The Robotics Revolution Is Finally Here

I've watched a decade of CES robot demos follow the same script: a company wheels out a machine that folds a towel or pours a drink, the crowd claps, and by spring nobody's heard from it again. So I went into this year's "physical AI" push assuming it was the same story with a rebranded label. Some of it was. But enough of it wasn't that I think it's worth writing down, plainly, what I actually think changed and what didn't.

What "physical AI" actually means

The phrase is doing real work, not just marketing. For most of robotics history, a robot's software stack was perception, planning, and control bolted together as separate hand-tuned modules — brittle, and every new task meant re-engineering a chunk of the pipeline. The shift this year is that a lot of teams are now building on vision-language-action (VLA) models: a single network trained on large amounts of video, language, and teleoperated demonstration data that outputs actions directly, the same way an LLM outputs tokens. NVIDIA's GR00T line, Physical Intelligence's Pi-0, and the RT-2/OpenVLA lineage out of Google DeepMind and the broader academic community are the clearest examples. Instead of writing a control policy from scratch for "pick up this specific bin," you fine-tune a generalist model on a few hundred demonstrations and it transfers reasonably well to a new task or even a new arm.

That's a genuine architectural change, not just better PR copy. It's the same pattern I watched play out with foundation models in NLP and vision — a slow multi-year build-up of pretraining infrastructure that suddenly makes fine-tuning cheap. Robotics is a few years behind, but it's now visibly on that curve.

The company landscape, and who I'd actually watch

Strip away the keynote lighting and the humanoid roster looks like this: Figure has a deployment agreement with BMW's Spartanburg plant; Apptronik's Apollo is doing similar pilot work with Mercedes-Benz and GXO logistics; Agility Robotics' Digit has quietly been moving totes in warehouses for a couple of years now and is, by most accounts, the closest thing to an actual production humanoid today, precisely because its job is boring; Tesla keeps showing Optimus doing increasingly polished tasks but hasn't shipped it as a product anyone can buy or rent; 1X is betting on a home-and-light-commercial humanoid (Neo); and Unitree's much cheaper hardware out of China is doing to humanoid bodies what budget quadcopters did to the drone market — dragging the price floor down fast, which matters more for the field's trajectory than any single flagship demo.

None of that is hype-free. A demo on a stage, run on a curated task with an operator a few feet away holding a kill switch, is not the same claim as "this robot restocks shelves at your store." I'd read every one of these announcements the way I read a startup's Series A press release: real signal about direction and funding, weak signal about what ships to a customer next quarter.

Where I'm skeptical: the humanoid form factor itself

Here's my actual contrarian take, for what it's worth: I don't think the humanoid shape is obviously the right answer for most of the jobs currently being pitched, and I think part of its popularity is that a two-legged robot photographs well and raises money more easily than "a robot arm bolted to a wheeled cart," even when the cart is cheaper, more reliable, and does the job just as well. Bipedal locomotion adds a huge amount of mechanical and control complexity — balance, gait planning, fall recovery — for tasks that a stable wheeled base handles trivially. The argument for humanoids is that the built environment (stairs, door handles, shelving heights) is designed for human bodies, so a human-shaped robot needs no retrofitting. That's a real argument, and it's why I don't dismiss the category outright. But for warehouse and light-manufacturing work specifically — which is where nearly all the current pilots are — wheels and a manipulator arm cover most of it today, cheaper and sooner.

The technical constraints that actually gate progress

A few things are genuinely hard right now, and I don't think they get resolved by a keynote:

• Sim-to-real transfer is still lossy. Domain randomization in Isaac Sim, MuJoCo, or Genesis gets you a policy that mostly works, but contact-rich manipulation — anything involving friction, deformable materials, or precise force control — degrades noticeably moving from simulation to a physical gripper. This is an old problem and it's improved, not solved.
• Inference latency versus model size is a real tradeoff. A multi-billion-parameter VLA model is too slow to run as a tight reactive control loop on edge hardware. The common workaround — a large "slow" model for high-level planning paired with a small, fast reactive controller for moment-to-moment control — is a reasonable engineering compromise, not evidence the problem went away.
• Actuators and tactile sensing are still expensive. Precision harmonic-drive actuators remain the single biggest line item in a humanoid's bill of materials, and cheaper alternatives generally sacrifice the torque control fidelity you need for safe contact with people or fragile objects. Tactile sensors (GelSight-style optical sensors, capacitive skins) have gotten meaningfully better, but adding them is still a real cost and integration decision, not a checkbox.
• The whole approach is data-hungry, and data collection is a bottleneck only well-funded players can push through at scale. Teleoperated demonstration data is the fuel for these generalist policies, and running a large enough teleop fleet to matter is expensive. That's a structural reason the field is currently shaped by whoever has the most capital, not necessarily whoever has the best idea.

What I'd actually pay attention to

If I were picking where to spend my own time, it wouldn't be chasing the humanoid narrative. It would be the boring middle: fixed or wheeled arms doing well-defined manipulation in warehouses and light manufacturing, using an off-the-shelf VLA backbone fine-tuned on a modest, well-collected demonstration set. That's where the unit economics already work, where Isaac Sim and similar free-to-cheap simulators let a small team iterate without a robotics lab budget, and where the gap between demo and deployment is smallest. The humanoids will keep getting headlines, and some of that progress is real. I'm just not convinced "human-shaped" is the detail that matters most yet — and I'd rather be wrong betting on the boring thing than wrong betting on the exciting one.