Agility Digit 5 Humanoid Kneels Near Workers to Cut Injury Risk

A humanoid that moves out of the way first

Agility Robotics has introduced Digit 5, a new version of its general-purpose humanoid. It is the first Digit built for what the company calls "cooperatively safe" work, meaning it can operate near people without the physical barriers that usually fence off industrial automation.18 The feature getting the most attention is how it reacts to people. If it spots someone at a distance, Digit 5 can steer around them or hold still and let them pass. If a person gets closer, it can squat down into a seated position.2 Agility describes the system as proprietary AI algorithms running on several kinds of sensors. They watch for people continuously and choose to avoid, stop or sit, while visual and audio cues tell nearby workers what the robot is about to do.18

The robot is not shipping yet. Agility expects early access to start in the first half of 2027 and general availability by the end of that year, and it is retooling its RoboFab plant in Salem, Oregon, to build the new model.11 Still, the launch is one of the clearest signs so far that the race to put humanoids in workplaces now depends on safety at least as much as on agility or dexterity.

How the coverage framed it

The reporting agrees on the basic facts and differs mostly in tone. Futurism said the robot "cowers in fear" and drops to its knees when a person comes around a corner in Agility's promotional video.15 The Cool Down took up the same framing and pointed out that the design puts the burden of adjusting on the robot, not the worker.12 The technical coverage was more restrained. Ars Technica quoted CTO Pras Velagapudi describing a "complex safe motion system" that picks different responses depending on what kind of human presence it detects.2 IEEE Spectrum gave the most useful explanation of why the robot kneels: the aim is to make it physically impossible for the 284-pound robot to fall on anyone. As a person approaches, Digit puts down its load and settles into a stable seated position before they reach it.1918

The Spectrum explanation is the right one. The kneeling is not meant to look polite. It addresses the specific hazard of a bipedal robot, which can tip over, and treats a toppling robot as the main risk to manage. That ties into the standards work. ISO 25785-1, the draft standard for dynamically stable robots, is meant to quantify how far and in which direction a walking robot might fall, which has been a major gap in existing rules.1

Some coverage went too far. One outlet said Digit 5 can work alongside people with "no risk" of serious injury.10 Neither Agility nor the more careful reporting claims that. Velagapudi presented the current "cooperative safety" mode as a first step. In this mode, robots share space with people but keep a safe distance. Only after it is proven in commercial use would the company move to "collaborative safety," where humanoids work right next to people.2

The injury argument

There are two separate injury questions here, and the launch speaks to both. The first is whether the robot itself hurts people. The industry has reason to worry. A compliance guide cites a reported February 2025 case at Tesla's Fremont plant in which a worker was knocked unconscious and pinned by a robotic arm during maintenance. It also cites research finding that unexpected activation causes most robot-related workplace accidents.1 As of mid-2026, OSHA had not issued guidance specific to humanoids, so employers are judged under the agency's General Duty Clause.3

The second question, which gets less attention, is how many injuries the robot prevents by taking over strenuous work. Agility says a redesigned leg with proprietary cycloidal actuators lets Digit 5 repeatedly lift 50 pounds, 40% more than before. That covers every single-person lift task in OSHA-regulated facilities.18 Reach also rises from 5.5 feet on Digit 4 to 7.2 feet, enough for shelves built for adults of average height.11 This matters because Toyota's reasons for deploying earlier Digits in Canada reportedly had less to do with replacing workers than with cutting ergonomic injury risk from repetitive walking and lifting.3 Researchers studying humanoids in construction make a similar case. They note that manual material handling is a leading cause of musculoskeletal disorders.4

The two questions are linked. A robot that can only work safely inside a cage can't take over lifting in the shared aisles where people actually do it. Velagapudi said as much: palletizing while a worker swaps out pallets, or moving down aisles where people are also working, are jobs where caging the robot isn't practical.2

The robot foundation model angle

For anyone following robot foundation models, Digit 5's most important feature may be its compute and safety design, not its kneeling. The robot runs Nvidia's IGX Thor module and is the first launch partner for Nvidia's Halos for Robotics platform.18 Nvidia expanded Halos from self-driving cars to robots in June 2026. IGX Thor includes a separate processor dedicated to safety workloads, so the functional system and the safety system run on the same chip.17 Agility's own description is similar: an independent safety controller monitors how Digit responds when someone comes within an unsafe distance and triggers the right reaction.18

This split matters because learned models are probabilistic. A robotics explainer notes that a hallucinated action can mean a dropped object or a safety hazard, not just a wrong sentence, and that strong safety guarantees for foundation models are still an unsolved problem.9 The way the industry is responding, as reflected in Digit 5 and Halos, is to keep generalist AI from being the last line of defense. Learned policies can make the robot more capable, while a separate, verifiable controller handles the decisions that protect people. That reading is my analysis, but it matches what Nvidia's Goel told Ars: robots with generalized bodies and generalized skills can only work across many tasks and settings in production if safety improves.17

The capability side is moving quickly. Nvidia has released Isaac GR00T N1.7 for general use. It is an open, commercially licensed vision-language-action model for humanoid skills, pretrained on robot data plus 20,000 hours of human video, and it supports whole-body control that coordinates manipulation and locomotion.7 Agility, for its part, says Digit 5 will use its "proprietary approach to physical AI-based capability training" to add manipulation skills, so one robot could handle a facility's whole flow, from depalletizing through machine tending, kitting, inspection and palletizing.18 One industry tracker points out that Agility has published few details about its models, which makes it hard to compare how much of Digit's behavior comes from learned policies.8

What remains unproven

The most important caveat comes from that same availability review: Digit's production record belongs to Digit 4, not Digit 5.8 That record is real. Earlier Digits have logged more than 65,000 hours at customers including GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada.11 At GXO's site near Atlanta, Digit 4 handled 100,000 totes at about 98% accuracy while on task.18 But a robot that sits down whenever someone approaches will lose throughput in busy facilities. No coverage so far measures how often that would happen in practice.

The cost numbers need the same caution. IEEE Spectrum estimates that Digit 5 sold as a service could save an employer about $100,000 a year per robot. That figure assumes 20 hours of work a day and a fully loaded labor cost of $30.50 an hour.19 The 20-hour claim depends on a 90-minute battery that recharges in nine minutes, a 10:1 run-to-charge ratio compared with Digit 4's 2:1.18 Both numbers are reasonable, and both still need to hold up in customer sites.

The rules are also unsettled. Agility is helping write ANSI/A3 TR R15.108 for dynamically stable industrial mobile robots in the US and Canada, and it contributes to ISO 25785-1.18 The ISO draft is still under committee review before a vote by the organization's 89 voting member nations.2 Helping shape the standard gives Agility an advantage. It also means the safety case for Digit 5 is being written alongside the rules it will be judged against.

The takeaway

The coverage presents Digit 5 as a robot that "adapts" to people. More precisely, it is a robot designed so that its learned skills can grow without the most dangerous decisions depending on those skills. Hurst, Agility's co-founder, has called safety a "blocker for everybody" in the humanoid industry.2 Digit 5 doesn't remove that obstacle yet, but it gives a believable outline of how it could be removed: independent safety hardware, a stable fallback posture and standards that measure falls. If the 2027 deployments show the robot can stay productive while giving way to people, other humanoid makers will probably have to copy that design or explain why they didn't.