Agility's New Humanoid Sits Down When You Get Too Close

On September 15, Agility Robotics unveiled Digit 5, and the headline feature is not the payload, the actuators, or the battery. It is a separate controller whose only job is to decide that a human got too close.

When it decides that, Digit 5 can avoid, stop, or sit down. Autonomously. Without the policy stack that is doing the actual work getting a vote.

The spec sheet

Digit 5 stands 5 feet 11 inches, weighs 284 pounds, and can reach up to 7.2 feet. Against the previous generation it carries 40% more payload, rated for repeated lifts to 50 pounds.

The legs were redesigned around cycloidal actuators. End-effectors are swappable and ISO-standard, which means the gripper is a procurement decision rather than an engineering project. The battery runs 90 minutes and recharges in nine — a 10:1 run-to-charge ratio, which is the number that determines how many robots you buy to cover one shift.

Agility reports more than $300 million in multi-year orders, with a pipeline across manufacturing, warehousing and logistics.

Why the safety controller is the product

Industrial automation has always been fenced. A robot arm that can move fast enough to be useful is a robot arm that can move fast enough to kill you, so it lives inside a cage, behind a light curtain, in a cell that shuts down when the perimeter breaks.

That architecture has a hidden cost that nobody puts on the quote: floor space, and the fact that the cell cannot be reconfigured casually. Every fence is a commitment to a layout.

Agility's framing for Digit 5 is "cooperatively safe work at scale" — a humanoid engineered to operate in close proximity to people without the physical barriers traditional automation requires. The mechanism is a safety architecture with proprietary AI, multiple sensors running continuous human detection, visual and auditory cues so people know what the robot is about to do, and an independent safety controller that acts on its own authority.

The independence is the engineering claim that matters. A learned policy deciding it is safe to keep moving is a probabilistic statement. A separate, simpler controller with veto power over the motors is the kind of thing a safety assessor can actually reason about, because its behavior does not change when you update the model.

Sitting down as a failure mode is a nice touch and not a cosmetic one. A 284-pound biped that stops in place is still a 284-pound biped standing over you. One that assumes a seated position has lowered its center of mass and taken tipping off the table.

What $300 million in orders does and does not tell you

It tells you the buyers are real and the contracts are multi-year. It does not tell you how many robots that is, over what period, with what cancellation terms, or how much is contingent on deployment milestones.

Agility has been more disciplined than most of this sector about reporting operational metrics rather than unit counts — the company's prior disclosures have leaned on hours worked, which is the honest measure when the question is whether a robot is actually doing a job. An order book is a different kind of number. It is a forward statement, and it arrives as the company moves toward a public listing.

Read it as demand signal, not as revenue. The useful follow-up is the same one that applies to every humanoid announcement this year: how many hours per week, at how many sites, doing what, with how much remote intervention.

The competitive frame

Global humanoid shipments in the first half of 2026 ran roughly 19,000–22,000 units, up about 272% year over year, with Chinese manufacturers holding something like 93–97% of that volume. Those are mostly not warehouse workers — a large share is companion, education and demonstration hardware, sold into a domestic market that is subsidizing the category aggressively.

Agility is not competing in that volume fight and has not tried to. Digit 5 is priced and engineered for a customer who is comparing it against a conveyor retrofit and a headcount line, not against a consumer robot.

That positioning has a defensible moat and a hard ceiling. The moat is certification: safety architecture that satisfies an industrial buyer's insurer is slow to build and slow to copy. The ceiling is that the addressable market is the set of tasks where a bipedal form factor beats a wheeled base with an arm — which is a narrower set than the marketing for this entire industry implies.

Cycloidal actuators and a nine-minute recharge are both signals that Agility knows which side of that line it is on. Those are duty-cycle investments. You make them when you expect the machine to work repetitive shifts, not when you expect it to do demos.

What to watch

Whether the safety controller gets certified, and to what standard. "Cooperatively safe" is a marketing phrase until a notified body attaches a number to it. A published conformity assessment would be the single most valuable disclosure Agility could make.

Whether the ISO end-effector standard gets adopted by anyone else. Swappable tooling only becomes an ecosystem if a second manufacturer honors the interface. If it stays proprietary-in-practice, it is a convenience, not a platform.

The intervention rate. Every humanoid in a real facility has a human somewhere in the loop. The number that separates a product from a pilot is how often that human is needed — and nobody in this sector publishes it.