Hello Digit 5: Agility Robotics releases a key candidate for a safe robot coworker

Digit 5: what Agility announced

Agility Robotics unveiled Digit 5 on September 15, 2026, from its headquarters in Salem, Oregon. It is the fifth generation of Digit and, by the company's framing, the first humanoid engineered for "cooperatively safe work at scale" ... designed to operate in close proximity to people without the physical safety barriers that traditional industrial automation requires.

What it is built to do

Digit 4, in commercial service since late 2023, has largely done one thing: tote-based material handling. 

Agility is pitching Digit 5 as a single robot doing much more: covering a facility's full workflow including depalletizing inbound goods, machine tending, kitting and sequencing, quality inspection, and palletizing outbound shipments. 

Agility attributes the expansion to its physical AI capability-training stack, and says Digit coordinates handoffs with a facility's existing AMRs, conveyors and warehouse systems through Agility Arc, its cloud fleet platform, which connects to WMS, WES and MES software and reports fleet KPIs including uptime, throughput and mean time between incidents.

The physical specification

Digit 5 is BIG.

It weighs 284 lb (129 kg) and stands 5' 11" (1.81 m). It lifts up to 50 lb (22.7 kg) repeatedly — 40% more than Digit 4 — via a new leg design built around proprietary cycloidal actuator technology, which Agility says covers the full range of single-person lift tasks in OSHA-regulated facilities. 

Digit 5's maximum reach is 7.2 ft (2.2 m), up from Digit 4's 5.5 ft, while keeping a footprint sized for human aisles, doorways and workstations. Grippers are now swappable on ISO-standard mounting flanges.

Power and duty cycle

The new battery system runs 90 minutes, which is down from Digit 4's 4 hours, but rapidly recharges in just 9 minutes, a 10:1 run-to-charge ratio against Digit 4's 2:1.

Agility claims that yields more than 20 hours of productive work in a 24-hour day. If that pans out in real working scenarios, that's 2.5 human shifts a day.

Charging is autonomous: Digit walks to the charge station, backs in, and recharges.

The safety architecture

There are 3 disclosed functions:

1. Safe human detection uses proprietary AI algorithms across multiple sensor technologies to monitor continuously for people, responding by autonomously avoiding, stopping, or assuming a seated position. 
2. Safety cues convey motion intent through visual and auditory signals. 
3. Safe motion control puts an independent safety controller in charge of the response when a person is detected inside an unsafe distance. Agility is the first launch partner for NVIDIA's Halos for Robotics platform, pairing Digit with NVIDIA IGX Thor and Halos Core, and will use NVIDIA's Halos AI Systems Inspection Lab to shorten its path to certification. The company is also a project leader on ANSI/A3 TR R15.108 and a contributor to ISO 25785-1, the first international safety standard scoped to the humanoid category.

The track record behind it

Digit has the longest tenure of any humanoid actively working in industrial facilities, thanks to being the first humanoid robot to get a paying gig.

(I covered that story way back in 2024.)

So Agility Robotics has nearly three years of continuous commercial deployment including more than 65,000 logged operational hours across customer sites such as GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada. 

At GXO's Flowery Branch facility near Atlanta, Digit 4 passed a cumulative 100,000-tote milestone in November 2025 at approximately 98% accuracy while on task. In November 2025 Digit also became the first humanoid to pass an independent field evaluation on a customer production line against industrial safety standards administered through OSHA.

The commercial position

As of May 2026, Agility reported more than $300 million in multi-year customer orders for Digit 5, with a pipeline across manufacturing, warehousing and logistics.

"Digit 5 is removing a major barrier to scaling humanoid robots in industrial environments," said Peggy Johnson, CEO of Agility Robotics. "We built it to the exact requirements our customers gave us after three years of Digit 4 working on their production floors, and the market has responded with more than $300 million in multi-year orders and a growing pipeline of customers across manufacturing, warehousing, and logistics."

Where it gets built, and when you can get one

Digit is assembled at RoboFab, Agility's 70,000-square-foot Salem facility. It's designed to reach 10,000 robots per year and will have 500-plus employees at full capacity. (Plus how many working humanoid robots, I wonder.) Agility's Pittsburgh location handles engineering testing, validationing, and skills development, and a Fremont, California site that opened in July 2026 as the company's hardware and physical AI development hub. 

Early access to Digit 5 begins in the first half of 2027, with general availability for manufacturing, warehouse and distribution operators expected by the end of 2027. 

Next year, for the first time, Digit will be also besold outside North America, starting with the EU and UK, with a CE mark and additional regulatory marks expected.

The analysis that follows

Digit 5 is the first humanoid engineered from the outset around a certifiable functional safety architecture, and that is a meaningful first. 

The design choices that make it possible — a dedicated safety island, redundant human detection, a statically stable degraded pose, and cycloidal actuators sized for fatigue life rather than peak torque — do cost the platform roughly 65 kg of mass and more than half its per-charge endurance, as far as I can tell.

Here's what I'm seeing ...

1. 129 kg, almost 200 pounds

The weight is kind of a big deal for a human-safe robot.

There are two established ways to let a robot share space with a human. Power and force limiting (PFL) caps the energy the machine can deliver into a body, using the biomechanical limit tables in ISO/TS 15066. On the other hand, speed and separation monitoring (SSM) lets the machine carry as much energy as it likes, and guarantees it never touches anyone: the robot continuously computes a protective separation distance and modulates speed or stops before that distance is breached.

At this point a 129 kg bipedal robot with heavy lifting capability probably cannot pass PFL, even with  padding and compliant joints. At least not at any useful speed. The mass alone puts it outside the transient contact limits, most likely.

So Digit 5 is unambiguously an SSM machine.

You can see that in Agility's launch video: Digit 5 stops moving when humans even approaching its space: when they are pretty far away, in fact. And if humans get closer, Digit 5 goes into the robotic version of a fetal position: a super-low squat right down on the ground, clearly indicating it's not going to move and it's not going to fall on anyone.

Once you accept that designation, mass stops being a safety liability in the certification sense.

At this point the robot is not collaborative in the ISO 10218 / TS 15066 sense, or in the working alongside humans sense. You will not be handing Digit 5 a 2 x 4 and asking it to pass you the nails.

2. Safety architecture: a 2-channel decomposition

As mentioned above, Agility's 3 has multiple disclosed safety functions:

1. Safe human detection
   This uses proprietary algorithms across multiple sensor technologies. Multiple modalities is the operative detail ... diverse redundancy is how you get diagnostic coverage on a detection channel without a single common-cause failure mode taking out both. 
2. Safety cues
   This is visual and auditory indication of motion intent. it's not a safety function in the ISO 13849 sense — it does not reduce risk through control reliability — but it is a legitimate risk reduction measure under ISO 12100's hierarchy, and it is the layer most humanoid developers have skipped entirely. Forklifts and AMRs earned floor access partly on this.
3. Safe motion control
   This is an independent safety controller that oversees the response when a person is detected inside an unsafe distance. Independent is the load-bearing word, and this is where NVIDIA matters.

IGX Thor carries a dedicated Functional Safety Island rated IEC 61508 SIL 3 capable, with up to 12K DMIPs of compute, its own I/O, power and clock domains, and 22,000-plus safety mechanisms providing diagnostic coverage across the SoC. NVIDIA's own architecture documentation places the Safety Decision Maker — a finite state machine, not a neural network — on the FSI, physically isolated from the AI compute domain. 

That domain is substantial in its own right: up to 2,070 FP4 TFLOPS, 14 Neoverse ARM cores, 128 GB at 273 GB/s. (All of this cannot be cheap.)

In the Linux-plus-QNX configuration, an NVIDIA hypervisor partitions the platform further, running AI workloads in a Linux VM and safety-critical functions in a QNX VM.

The structural point: the perception stack proposes, and a deterministic state machine on separate silicon disposes. ISO 13849 Category 3 and 4 architectures require exactly this kind of channel separation, and it is genuinely hard to retrofit. 

Building on IGX Thor from the start is the most consequential engineering decision in Digit 5, more so than anything in the mechanical spec.

The Halos AI Systems Inspection Lab participation is a procedural shortcut rather than a technical one: NVIDIA's platform elements arrive pre-assessed, so Agility's certification scope narrows to application-specific logic instead of the whole stack. 

For a company trying to hit a 2027 date in two regulatory jurisdictions, that helps.

3. The leg problem: a robot with no inherently safe state

This is where humanoids diverge from pretty much everything the existing mobile robot standards were written for, and it explains why Digit 5's response set has three members instead of one.

For a wheeled AMR, the safe state is trivial: cut drive torque and engage brakes. The machine sits still. Safety-rated monitored stop is a solved problem, covered by ISO 3691-4 and ANSI/RIA R15.08.

A dynamically stable biped has no equivalent state. 

A Digit that stops actuating probably falls over, and a 129 kg mass falling from a 1.81 m standing height delivers a good chunk of potential energy into the floor or whoever is nearby.

Ouch.

Agility's disclosed responses are, in escalating order: autonomously avoid, stop, or assume a seated position. That third option is the on I mentioned above -- squatting -- and it is the most interesting one. 

Sitting is a controlled transition to a statically stable pose, lowering the center of mass, widening the support polygon, and removing the balance controller from the safety-critical path entirely. (And yes, vastly reducing that potential energy.) This looks to be a genuine safe state, reachable under degraded conditions, and it is the closest industry  an answer for "what does E-stop mean on a biped."

4. Digit 5 actuation is cycloidal, and the reason is fatigue, not force

Agility attributes the payload increase to "a new leg design engineered to withstand repetitive lifting, powered by proprietary cycloidal actuator technology."

Withstand repetitive lifting is a fatigue-life claim, not a peak-torque claim. That tells you what was missing in Digit 4.

Harmonic drives achieve high ratio and near-zero backlash through a flexspline that elastically deforms every revolution. Under repetitive shock, however, — heel strike, load pickup, sudden deceleration — the flexspline is a fatigue-limited element. Cycloidal drives, on the other hand, distribute load across many simultaneous contact points with far better tooth stress distribution. And that's why they tolerate large dynamic loads and frequent impacts without as much degradation.

This architecture decision, however, is not free:

• Backlash runs around 7 arcminutes in published cycloidal QDD designs, materially worse than a harmonic drive's near-zero. For locomotion and tote handling this is irrelevant. For the machine tending and quality inspection tasks Agility is now claiming, it gets more interesting.
• Torque ripple from the eccentric cycloid geometry is inherent: the referenced work measures roughly ±1.5 Nm amplitude and needs a learned estimator to compensate. Ripple degrades the fidelity of joint torque sensing, and joint torque sensing is one of the signals a functional safety argument for a legged robot typically leans on.
• Backdrivability is a genuine benefit: static and dynamic backdriving torques around 1.99 Nm and 1.36 Nm in the reference design mean the joint absorbs impact rather than transmitting it into the structure.

That last property is safety-relevant. A backdrivable joint is more forgiving in an unplanned contact, which partially offsets the mass penalty ... though not nearly enough to bring a 129 kg machine into PFL territory.

5. Mass budget: what an extra 65 kg buys

Digit 4 was probalby around 64 kilograms, although there's no 1 published 100% guaranteed source for that. That's a lot less than Digit 5.

Here's what the extra mass provides:

The ratio that matters most?

Roughly 9.7 kg of added robot per kg of added payload. That isn't an amazing exchange rate by any conventional mechanical standard, and it tells you the mass went somewhere other than lift capacity. Most likely that's in redundant sensing, safety compute, structural reinforcement for fatigue life, heavier cycloidal gearsets, and pack.

Payload fraction falling from about 25% to 17.6% is the trade-off. 

For external reference, Figure 03 claims 20 kg on a 61 kg frame, roughly 33%, though against a platform making no comparable safety certification claim. And clearly, Figure 03 is not an industrial robot platform in the sense that Digit 5 is obviously intended to be.

Digit 5 looks like a purpose-built factory robot. Figure looks more home, retail, service, with some capability in factories and warehouses.

(This is purely a visual/visceral argument, just to be clear.)

6. Energy: the runtime drop is probably a consequence

From Digit 4 to Digit 5, per-charge endurance fell from about four hours to 90 minutes, and the most likely explanation is not a smaller pack. 

Doubling mass roughly doubles the cost of transport. 

A 129 kg biped walking the same aisles at the same speed draws dramatically more power than a 64 kg one. The pack may be the same size or larger; the runtime collapse is probably a power-draw consequence of the mass budget.

Duty cycle, Digit 5: 90 min run + 9 min charge = 99 min cycle → 14.5 cycles per 24 h → ~21.8 h of run time. "More than 20 hours" is honest.

Duty cycle, Digit 4: 4 h run at 2:1 → 2 h charge → 6 h cycle → 4 cycles → ~16 h. The generational gain is real: +36% theoretical uptime.

Two operational consequences:

Dock transit overhead. Charge events go from about 4 per day to about 14.5. Every one requires walking to a dock and back. How long does that round-trip take?

Charging infrastructure. Replenishing a 90-minute pack in 9 minutes is a high C-rate event. If usable capacity is on the order of 1.5 kWh, that implies roughly 10 kW delivered per robot per charge, with the thermal management and cell chemistry to survive it 14.5 times a day, every day. For a 20-robot fleet you might have to look at a facility electrical upgrade.

7. Reach and the payload envelope question

Reach goes from 5.5 ft to 7.2 ft (1.68 m to 2.2 m), a 31% increase on a robot 1.81 m tall. 

Impressively, Digit 5 can now place a load well above its own head height.

For a biped this is a nontrivial stability problem. Extending a 22.7 kg mass to 2.2 m moves the center of mass up and forward, shrinking the stability margin exactly when the arm is at maximum moment arm. 

I do wonder if Digit 5 can lift its full 50-pound payload that high ...

8. Standards: shipping ahead of the rulebook

Digit 5's regulatory position is unusually well developed for this category and still rests on ground that has not set.

What exists and has been done. In November 2025 Digit became the first humanoid to pass a field evaluation by a Nationally Recognized Testing Laboratory — an OSHA-recognized certifier — on a live customer fulfillment line, assessed against ANSI/RIA R15.08, ISO 13849 and ISO 12100. Agility stated at the time that the evaluation was specific to that site's configuration and environment.

R15.08 is a three-part framework.

Part 1 covers the robot itself, Part 2 system integration and traffic management, Part 3 application-specific safeguarding. It applies an S1–S4 severity scale with exposure and avoidance scoring, and leans on ISO 13849 for the performance level architecture of safety functions. It was written for wheeled industrial mobile robots, though.

What does not exist yet. ANSI/A3 TR R15.108, covering dynamically stable industrial mobile robots in the US and Canada, is a technical report still under development, with Agility as a project leader. ISO 25785-1 — ISO/TC 299, scoped to industrial mobile robots with actively controlled stability, explicitly including legged platforms — is at Stage 30.60, committee draft with the comment period closed July 8, 2026. It still has DIS, FDIS and publication ahead of it. 2027 is optimistic; 2028 is realistic.

So general availability at the end of 2027 lands Digit 5 in customer facilities at approximately the same time as, or ahead of, the first international standard written for what it is.

The EU wrinkle. Early access opens in the first half of 2027 for the EU and UK, with a CE mark expected. EU Machinery Regulation 2023/1230 becomes applicable on 20 January 2027, replacing the Machinery Directive, with specific provisions for machinery embedding self-evolving behaviour and for safety components incorporating AI. 

Digit 5's European entry therefore coincides almost exactly with the new regime taking force. Whether that is good timing or a conformity assessment problem depends on notified body capacity and on how the AI-related provisions are read for a system like this.

10. Undisclosed parameters

For index completeness, Digit 5 has no published figure for: 

• walk speed (Digit 4 was approximately 1.5 m/s)
• degrees of freedom or joint count
• battery capacity in kWh
• dock power draw
• gripper DoF / grip force / finger count
• payload as a function of reach height
• IP rating
• operating temperature range
• acoustic emissions,
• actuator torque density or peak joint torque
• price and RaaS rate (though Agility's June 2026 investor deck modeled roughly $8,500 per robot per month)

I'm sure more of that will become available as Digit 5 nears wide availability in 2027. And most likely numbers published now would change between now and then.

Overall assessment

Digit 5 is the first humanoid where the safety architecture is clearly the primary design driver rather than a compliance layer added afterward, and the choice to build on a SIL 3 capable safety island with hardware-enforced separation between the AI and safety channels is the right one — structurally difficult to retrofit, and the reason Agility has a plausible path to certification while most of the field has a demo.

The seated safe state is the most interesting piece of safety engineering in the announcement, because it addresses the problem unique to this form factor: a dynamically stable machine has no passive safe state, and stopping is itself a hazard.

The costs are real, however.

A 102% mass increase for 40% more payload is the price of an SSM-certifiable platform, not a mechanical improvement. The endurance collapse from four hours to 90 minutes is probably a consequence of that mass, and it requires 10 additional dock transits per robot per day.

Overall: very impressive. But ... questions (as always) remain.