Scientists Put Digit 5 Next to a Human, and You Won’t BELIEVE What Happens!

Agility Robotics may have just changed the humanoid game again, and this time it is not about another viral robot demo. Its new Digit 5 is built around one problem factories cannot ignore, how a large humanoid can safely work around people. The robot can detect when someone gets too close, then avoid them, stop moving, or even lower itself into a stable seated position. It can repeatedly carry up to 50 pounds and squat much deeper than Digit 4. Its reach has jumped to more than 7 feet. Its battery only runs for around 90 minutes, but Agility says it can recharge in just 9 minutes. And instead of giving Digit 1 permanent hand, the robot can use different tools depending on the job. Even its famous backward-looking legs have been redesigned. So this is not just Digit 4 with a few upgrades. Agility has rebuilt the robot around safety, heavier work, and much longer daily operation. Hey guys, Alfie here. Welcome back to the AI Nexus. Digit 5 is Agility Robotics' 5th generation humanoid, and the company is positioning it very differently from many humanoids we usually cover. This is not being presented mainly as a research platform. Agility wants it working inside warehouses, distribution centers, factories, and logistics sites. The design came from roughly 3 years of feedback collected while Digit 4 was operating inside real customer facilities. Digit 4 already proved that a humanoid could do repetitive commercial work, but there was still a major restriction. The robot usually worked inside controlled areas that kept ordinary workers separated from it. Digit 5 is Agility's attempt to move beyond that model. Agility calls the idea cooperative safety, where dedicated safety hardware helps the robot operate closer to people. Digit 5 stands 5 feet 11 inches tall and weighs about 284 pounds. Its height is human-like, but it is much heavier than a person. Agility still wants it to fit through normal doors, move through human-sized aisles, and use workstations built for people. The first big hardware upgrade is payload. Digit 4 was rated for up to 35 pounds. Digit 5 raises that to 50 pounds for repeated material handling, roughly a 40% increase. This is not a one-time maximum lift. The robot is being designed to move those loads repeatedly during an industrial shift. The legs may be an even bigger change. Older Digit robots were known for their backward-looking leg design, which came from Agility's earlier Cassie robot. Those legs were excellent for dynamic walking, but they were not ideal when the robot needed to get very low to the floor. Digit 5 now uses a more human-like knee arrangement, and the goal is not simply to make the robot look more human. The new geometry lets Digit squat deeper and place its body in a better position when lifting from low shelves, pallets, or containers near the ground. Agility also built its own cycloidal actuators into the new design. These joints are meant to handle repeated heavy-duty motion. The focus is what happens after the robot has lifted the same kind of load hundreds or thousands of times. But the biggest headline is safety. Digit 5 includes what Agility calls safe human detection. Multiple sensors and a separate safety layer keep checking the area around the robot for people. If someone enters a risky distance, Digit does not have to rely only on its normal AI to decide what happens next. An independent safety controller can trigger a response immediately. Depending on the situation, Digit may avoid the person, stop moving, or lower itself into a seated position. That seated behavior is especially interesting because it places the robot into a more stable state instead of forcing a 284-pound humanoid to keep dynamically balancing while someone is close. Agility has also added visual and audio signals so people nearby can better understand what the robot is about to do. The company is also working with NVIDIA through its Halos for Robotics platform, using NVIDIA IGX-Thor and Halos Core as part of the wider safety architecture. And this safety work is not starting from zero. Agility says it is building on more than 65,000 hours of operating experience gathered from previous Digit deployments. If you want to keep up with humanoid robots that are moving from demos into real industrial work, subscribe to the AI Nexus, because the next change is one of the most unusual decisions Agility made with Digit 5. The battery actually looks worse at first. Digit 4 was advertised with around 4 hours of runtime, while Digit 5 is designed for about 90 minutes. But Agility completely changed the charging strategy. Digit 5 can recharge in around 9 minutes, giving it roughly a 10-to-1 operating-to-charging ratio. That means the robot does not need one huge battery session. It can work for about an hour and a half, recharge quickly, and then go back to work. Agility says this approach could give Digit 5 more than 20 productive hours during a 24-hour day. Charging is also autonomous. When the battery gets low, Digit is designed to return to its charger without waiting for an employee to plug it in. That becomes much more important if a warehouse eventually has dozens or even hundreds of these robots operating across several shifts. Reach has improved too. Digit 4 could reach roughly 5.5 feet vertically. Digit 5 can reach around 7.2 feet. That gives the robot access to much higher shelves and equipment without forcing companies to redesign everything around it. Agility has also moved away from the idea that one human-like hand should do every job. Digit 5 uses an ISO standard mounting point for interchangeable end effectors. In simple terms, different grippers or tools can be attached depending on the task. Agility has shown research involving five-finger robotic hands for industrial work. Agility is currently choosing simpler tools that can be strong and reliable instead of trying to copy the appearance of a human hand. And that connects directly to the biggest change in what Digit is supposed to do. Digit 4 became known mainly for tote movement. It could pick up containers, carry them between stations, work around conveyors, and place those containers somewhere else. That proved the basic idea, but it still meant the robot was doing a fairly narrow part of the overall workflow. Digit 5 is supposed to go much further. Agility lists depalletizing, machine tending, and palletizing as target jobs. Instead of buying a humanoid only to move the same tote between two points, the idea is that the same Digit platform could eventually perform several different jobs inside one facility. Machine tending could be especially important. Digit now has more reach, more payload, a deeper squat, and interchangeable tools. That combination could let the robot take material from a container, move toward a machine, load or remove a component, and then carry that material to the next part of the process. Behind all of this is Agility's physical AI stack. The company divides the intelligence into two broad areas. Semantic AI handles things like understanding a task, planning, and reasoning about what needs to happen. Physical AI deals with actually doing it. That includes walking, balancing, controlling force, manipulating objects, and coordinating the entire body. Agility also uses large language models and vision-language action models in that stack. So Digit 5 is not the first intelligent Digit. The bigger change is how that AI is being combined with new hardware, new safety systems, and a body designed for wider industrial use. And Agility has something many humanoid companies still do not have. A large amount of real deployment experience. Previous Digit robots have accumulated more than 65,000 operating hours. At GXO's Flowery Branch facility near Atlanta, Digit also passed 100,000 totes moved. Agility says the robot maintained around 98% accuracy while it was actively performing that work. Those deployments matter because Digit 5 was not designed only from simulation or lab testing. Agility says its training process combines real customer data, motion capture, teleoperation, simulation, and repeated practice. Digit 5 will also connect to Agility Arc, the company's fleet management platform. Arc can assign work, monitor performance, and connect Digit with warehouse software, conveyors, and autonomous mobile robots. That lets Digit become one part of a larger automated workflow instead of operating alone. But there is one important reality check here. Digit 5 has been unveiled, but it is not yet shipping at large scale. Agility expects early access during the first half of 2027, followed by broader availability by the end of that year. The company also plans to expand beyond the United States and Canada into Europe and the United Kingdom. Agility says it already has more than $300 million in multi-year Digit 5 orders, although those orders depend on product and contract milestones. Investor Material connects that commitment with roughly 1,000 robots under a three-year robots-as-a-service contract. So Agility has not already earned that entire $300 million, but customers are clearly planning around the new platform. Agility is also preparing for scale at RoboFab in Salem, Oregon. The humanoid factory is designed for peak output of up to 10,000 Digit robots per year, although actual production will depend on demand. And that is what makes Digit 5 different from a normal generational update. Digit 4 proved that a humanoid could keep doing a real industrial job for long periods. Digit 5 is Agility's attempt to take everything learned from that experience and remove some of the biggest limits around it. The robot can carry more, squat lower, reach higher, recharge far faster, switch tools, and cover more jobs. But the biggest change is still the safety architecture. If Agility can make this kind of cooperative human-robot work reliable in real facilities, Digit could spend far less of its working life behind barriers and far more of it moving through the same workplaces built for people. Digit 5 shows how humanoid hardware is becoming safer and more practical for real-world work. But Figure is attacking the next challenge from the intelligence side, with Helix 2.5 teaching robots how to work inside homes they have never seen before. Figure has just revealed Helix 2.5, its most advanced robot brain so far, and Figure built it around one surprisingly simple question. Can a humanoid robot enter a home it has never seen before and immediately start working on its own? That sounds like something robots should already be able to do, but it is actually one of the biggest problems standing between today's humanoids and truly useful home robots. A person can walk into someone else's house and still understand how to make the bed, fold a towel, or pick toys off the floor. The furniture can change. The room can be smaller. The objects can look different. We still understand what needs to be done. Robots normally don't work like that. And Helix 2.5 is Figure's attempt to change it. There are really three things happening with Helix 2.5 that are worth understanding. First, Figure's robot can now take behaviors it learned elsewhere and use them inside completely unfamiliar homes. Second, Figure has dramatically improved this ability by training the robot on huge amounts of human experience before teaching it individual chores. And third, Figure has now found evidence that simply scaling this human data could predictably make future versions of Helix better. That last point may eventually become more important than any of the demonstrations Figure showed this week. But first, we need to understand the problem. Helix 2.5 is trying to solve. Figure already made a major step with Helix-02. That model connected the robot's perception, walking, balance, and manipulation into whole-body autonomous behavior. Figure demonstrated it unloading a dishwasher and later running a logistics task autonomously over an extended deployment. But there was an important limitation. Those robots had learned using data collected in the environments where they were eventually expected to work. In other words, Figure could make the robot very capable inside a place it already knew. The harder problem was taking that behavior somewhere completely new. That difference becomes extremely important if Figure wants robots inside homes. A factory can be organized around a robot. The same workstation can stay in the same place. Boxes can arrive from predictable directions. But someone's house could have a bed pushed against a wall, toys underneath a table, different lighting, different towels, and furniture the robot has never encountered before. So Helix-2.5 was built to see whether the robot could carry what it had learned from one environment into another. Figure tested this using three chores. The first was tidying a living room. Between 13 and 15 toys were scattered around the room. And the robot had to collect every single one and place them into a basket. The second was towel folding. Every towel had to be folded and then placed into a basket. And the third was making a bed. The robot needed to correctly position both pillows, bring the comforter corners toward the top of the bed, and leave the comforter pulled smooth. These might sound like three fairly basic household chores. But together, they force nearly the entire robot to work as one coordinated unit. The robot has to see something, walk toward it, position its body, reach for it, manipulate it, and then continue to the next part of the task. With toys, it handles mostly rigid objects. With towels and bedding, it deals with material that bends, folds, changes shape, and moves every time the robot touches it. But Figure did not keep those behaviors inside one controlled test space. It took the robot into 30 real homes across the Bay Area and asked Helix 2.5 to perform them in places it had never worked before. And this is where the problem becomes much bigger than simply recognizing a toy or knowing how to fold a towel. A humanoid working inside a home has to control almost its entire body at once. It may need to walk across the room to find an object, move closer to get a better look at it, shift its feet into the right position, reach down, grab it, and then use both hands while staying balanced. And the objects themselves behave differently. A toy is mostly rigid. Once the robot grabs it, its shape stays the same. But towels and bedding constantly fold, sag, twist, and change shape as the robot touches them. So Helix cannot rely on one perfect sequence of movements. It has to continuously use what it sees to decide how its body should move next. Figure refers to part of this as active perception. Instead of expecting everything to already be visible from one position, the robot can move its body to get a better view, and then reposition again before interacting with something. That is especially important inside homes. Factories and robotics labs can be designed around the robot. A fixed robotic arm can have objects delivered directly into its workspace. A wheeled robot works best when there is enough clear floor around it. A home gives the robot none of those guarantees. Furniture blocks its path. Objects can be low to the floor or hidden behind something. A bed can be too large to reach from one side. The robot may simply be standing in the wrong place to finish the next part of the task. This is where the humanoid shape becomes useful. The same legs that bring the robot across the room can also reposition its entire body for manipulation. Walking is no longer just a way of traveling from one point to another. It becomes part of completing the task itself. And Helix 2.5 was specifically built around this idea of zero-shot whole-body autonomy. Figure calls it the most advanced neural network it has built so far. But one of the biggest changes is actually how that intelligence is created. The jump from Helix O2 to Helix 2.5 becomes even clearer in the success rates. In zero-shot testing across new environments, Figure O3 reached a 62% success rate in towel folding, 40% in toy tidying, and 67% in bed making. With Helix O2, those same tasks were only around 9%, 5%, and 11%. So Helix 2.5 did not just help the robot work in unfamiliar homes. It produced a massive improvement in how reliably those learned behaviors transferred to new environments. Unlike Helix O2, which started from an existing vision language model, Helix 2.5 was pre-trained from scratch using INDEX, Figure's global platform built to collect high-quality human behavior data from the real world. That broader training shows up in how the robot handles unfamiliar situations. While making a bed, for example, Helix can step back, change its stance, or walk around the bed when its current position is not enough to finish the task. Instead of blindly repeating a motion, the robot can reposition itself and continue. This is part of what Figure means by zero-shot generalization. Helix 2.5 can take a behavior learned elsewhere and perform it inside homes it has never seen before. Figure demonstrated this across 30 different homes, where the layout, furniture, and objects could change from one environment to another. The robot still has to understand what is around it, position its body correctly, and complete the same behavior despite those changes. So how did Figure get Helix to generalize this well? This is where INDEX becomes important. INDEX is Figure's global collection of recorded human physical behavior. Instead of relying only on robot demonstrations, Figure uses this human data to expose Helix to a much broader range of movement, objects, environments, and physical interactions before teaching it a specific robot task. The dataset is deliberately broad. Figure says no individual evaluation task made up more than 1.90% of INDEX pre-training data. So Helix is not simply watching huge amounts of bed-making or towel-folding. The goal is to build broader physical knowledge that can later transfer into different robot behaviors. INDEX is also growing quickly. Figure says it is now collecting roughly 35 minutes of new human experience every second. But Figure also tested whether this pre-training actually makes a difference. It trained two policies using the same task-specific robot data. The architecture, training setup, and evaluation were kept the same. The key difference was that one started without INDEX pre-training, while the other started from the INDEX pre-trained Helix 2.5 model. Without INDEX, the robot achieved a 9% zero-shot success rate. With INDEX, that increased to 56%. That is more than a six-fold improvement. Figure also found that Helix 2.5 needed about half as much task-specific data as a comparable Helix-02 behavior. Helix-02 had been trained directly inside the environment where it would later operate. Helix 2.5 used less task-specific data, matched that earlier success rate, and then transferred the behavior across 30 unseen homes. This is important for scaling humanoid robots. If every new home, factory, or warehouse requires its own large training dataset, deploying robots across thousands of locations becomes slow and expensive. Figure is trying to reverse that process. Give Helix broad physical knowledge first, use a smaller amount of data to teach a specific behavior, and then carry that behavior into new environments. That is also why Figure is investing so heavily in data and computing power. The company has committed $3.5 billion in compute toward training Helix, while INDEX continues collecting more human experience. The bigger goal is not simply teaching Figure's robot more individual tricks. It is building a foundation that needs less specialized training every time the robot encounters something new. And that is why Figure's message after Helix 2.5 is simple. It is time to scale up.