Agility's Digit 5 drops the safety cage — Europe's rules arrive in 2027

What Agility actually announced

According to the announcement covered by Unite.AI, Digit 5 is the fifth generation of Agility's general-purpose humanoid, and the central claim is "cooperatively safe" operation: the robot shares floor space with people instead of being fenced off from them.

The safety architecture has four visible layers. Proprietary AI collision-avoidance algorithms handle perception and prediction. A multi-sensor suite monitors the surroundings continuously. The robot itself emits visual and auditory cues so a human can read its intent before it moves. And an independent safety controller sits apart from the main compute, able to stop the machine or park it in a sitting position when a defined distance is breached.

Digit 5 is also the launch partner product for NVIDIA's Halos for Robotics platform, announced in June 2026, running on NVIDIA IGX Thor hardware with the Halos Core infrastructure underneath. That matters more than the marketing framing suggests: IGX Thor is the functional-safety-oriented edge tier in NVIDIA's lineup, and Halos is the layer that is supposed to make safety certification of autonomous machines tractable.

Where the scepticism belongs

Separating a certified safety controller from an AI perception stack is the right architectural instinct. The safety logic that stops the machine lives in hardware and firmware that can be validated; the probabilistic model that answers "is that a person or a wrapped pallet at eleven metres?" lives elsewhere.

The uncomfortable question is what happens when the AI layer is wrong in the other direction — when it decides a human is a pallet. A cage does not care how confident a neural network is. A distance-based stop triggered by an independent controller is only as good as the sensor data it receives, and if that data arrives from the same perception pipeline that the AI uses to plan its path, then the "independent" controller is independent in name more than in function.

Agility has 65,000+ hours of commercial deployments behind the platform, which is a real number and not a demo reel. Whether it is enough to convince a European notified body is a separate conversation, and it has not started yet.

The numbers, including the one that doesn't fit

Here is everything concrete that came out of the launch, compiled in one place:

Run the arithmetic and the shape of the story changes. A $2.5 billion valuation against $1.8 million in reported net sales is roughly 1,389 times revenue — the kind of multiple that only works if the $300 million order book converts. Those are multi-year orders, not invoices. Agility also has not published a unit price, and no European per-hour or per-unit rate was announced alongside Digit 5. There is no EUR figure to compare, because there is no figure at all.

For scale: on our own AI Arena benchmarking rig we can put a local model through its paces on an RTX 5060 Ti in an afternoon and publish real tokens-per-second numbers. You cannot benchmark a biped with a 22.7 kg payload on a workbench, which is precisely why robotics journalism defaults to press releases.

Europe is the difficult market, and not for the reason you think

Agility confirmed plans to expand into the European Union and the United Kingdom. No dates, no pricing, no launch countries. That vagueness is understandable once you look at the calendar.

The Machinery Regulation (EU) 2023/1230 replaces the old Machinery Directive and applies from 20 January 2027 — the same window in which Digit 5 reaches early access. The regulation explicitly anticipates machinery incorporating AI systems that perform safety functions, and requires that the safety of autonomous machines be maintained across the product lifecycle. A cage-free humanoid with an AI collision-avoidance layer is close to the textbook case, and it needs CE marking through a conformity assessment before it can legally be placed on the EU market.

Then there is the AI Act. Under Annex I, an AI system functioning as a safety component of a product covered by listed harmonisation legislation — machinery among them — and requiring third-party conformity assessment is classified high-risk. That triggers documentation, logging and human-oversight obligations on top of the machinery file. Two compliance tracks, one product, no announced date for either.

The point most people miss is GDPR. A 20-hour-per-day humanoid carrying cameras and depth sensors through a warehouse is a mobile personal-data processing platform. It will see faces, badges, gait, work pace. Before any European deployment, an operator needs a data protection impact assessment, a lawful basis, and answers about where telemetry and video are stored — including whether any of it leaves for US-hosted infrastructure.

The competition does not wait in Brussels

Europe is not a greenfield. Apptronik has publicly run Apollo pilots with Mercedes-Benz in Hungary. Neura Robotics, based in Metzingen, Germany, is building humanoids on European soil with a sovereignty argument attached, and PAL Robotics in Barcelona has been in service robotics for years. Chinese vendors keep pushing unit prices downward, which will matter more to a Czech logistics operator than any technical comparison sheet.

Agility's advantage is deployment hours and a partner in NVIDIA whose safety platform becomes a de facto standard if it works. Agility's problem is that "cooperatively safe" is a claim, and in the EU it becomes a certificate or it becomes nothing.

What a European operator should do now

Nothing on the Digit 5 timeline is binding for Europe. A warehouse manager in Czechia or Germany should treat early access (1H 2027) as a US-first programme, budget for a conformity assessment that has not been scoped, and start the DPIA long before the robot arrives — because the data protection question will be asked by the works council earlier than by the regulator.