Video Friday: Digit Learns to Dead-lift
Agility Robotics' Digit humanoid robot has learned to deadlift 65 pounds using a machine learning policy trained in simulation. This milestone shows that commercial humanoid robots are getting close to handling the physical demands of real warehouse work, which is the whole point...
According to IEEE Spectrum's robotics coverage, Agility Robotics has demonstrated that its Digit humanoid robot can now perform a full deadlift of 65 pounds (29.5 kilograms). The story was featured in the publication's recurring Video Friday series, which the IEEE Spectrum robotics team uses to spotlight the most compelling new developments from research labs and commercial robot developers each week. The clip itself is short, but the technical work behind it is not.
Why This Matters
A 65-pound deadlift sounds like a gym milestone, but for a bipedal robot operating outside a lab, it is a serious engineering achievement that directly tests whether humanoid robots can replace human workers in logistics and warehousing. Digit stands roughly 5 feet 4 inches tall and weighs around 125 pounds, making it physically comparable to a human worker, and that size only matters commercially if the robot can actually lift what humans lift on a warehouse floor. Agility Robotics is already selling Digit through a Robot-as-a-Service model, meaning this capability improvement directly expands the contract value they can offer customers. The companies still treating "humanoid robots in warehouses" as a 2030 problem should probably move that date .
Daily briefing from 50+ sources. Free, 5-minute read.
The Full Story
Agility Robotics built Digit to work alongside humans in warehouses and logistics environments, not to sit in a research lab. The robot is bipedal, roughly human-sized, and designed with arms and hands capable of grasping and manipulating objects. Until recently, the heavier end of object manipulation was a limitation. A 65-pound deadlift changes that calculus considerably.
The technical approach Agility used to train this capability centers on simulation-based machine learning. Rather than programming a rigid set of lifting instructions, the team trained a control policy by running Digit through thousands of virtual lifting repetitions. Crucially, they included the actual object being lifted inside the simulation environment during training, not just the robot itself. That detail matters because it forced the model to account for how a 65-pound load shifts the robot's center of mass, how grip forces change throughout the movement, and how load distribution affects balance at every point in the lift.
The persistent problem in robotics is what researchers call the sim-to-real gap: a robot that performs perfectly in simulation often fails the moment it touches actual hardware in the physical world. Physics engines are approximations, not perfect replicas of reality. Agility's approach of embedding the object's properties directly into the simulation produced a trained policy that successfully transferred to the real Digit hardware. The robot did not just approximate the deadlift; it completed it with the whole-body coordination the movement demands.
That whole-body coordination piece is worth slowing down on. A deadlift is not an arm exercise. It requires the legs, hips, spine, and arms to work as a coordinated system while an external load constantly tries to pull the body forward and off-balance. For a bipedal robot, maintaining dynamic balance under those conditions while managing actuator stress across multiple joints simultaneously is genuinely hard. Getting the policy to handle all of that in real time, on real hardware, represents progress that spreadsheets about actuator specs alone cannot capture.
Agility Robotics is also at a commercially interesting moment. The company sells access to Digit through a Robot-as-a-Service model, meaning customers pay for the robot's work rather than buying the hardware outright. Every new capability Digit demonstrates expands the range of tasks the company can contract for, which directly affects revenue potential. Demonstrating a 65-pound deadlift in April 2026 is not just an engineering flex; it is a sales pitch delivered on video.
The IEEE Spectrum feature placed this development alongside a calendar of upcoming robotics conferences, including the International Conference on Robotics and Automation (ICRA 2026) running June 1 through 5 in Vienna, Austria, and the Robotics: Science and Systems conference (RSS 2026) scheduled for July 13 through 17 in Sydney, Australia. Both events will likely feature further research on sim-to-real transfer and whole-body control policies, the two technical threads running through Digit's deadlift achievement.
Key Details
- Digit can now deadlift 65 pounds (29.5 kilograms) using a trained machine learning policy.
- The training methodology placed the 65-pound object directly inside the simulation environment to capture realistic load dynamics.
- Digit stands approximately 5 feet 4 inches tall and weighs around 125 pounds.
- Agility Robotics sells Digit access through a Robot-as-a-Service (RaaS) model.
- The achievement was featured in IEEE Spectrum's Video Friday series in April 2026.
- ICRA 2026 is scheduled for June 1 through 5, 2026 in Vienna, Austria.
- RSS 2026 is scheduled for July 13 through 17, 2026 in Sydney, Australia.
What's Next
Agility Robotics will almost certainly push the weight ceiling higher now that the simulation training methodology has proven it can transfer reliably to real hardware. Watch for Digit deployments in active warehouse contracts to start including heavier object handling specifications as Agility updates its RaaS offering in the second half of 2026. The ICRA and RSS conferences this summer will also be worth monitoring for peer-reviewed research on the specific policy training approach Agility used, since other humanoid robot developers will be studying this methodology closely.
How This Compares
Boston Dynamics has spent years demonstrating Atlas performing backflips and parkour routines, but those demonstrations have always prioritized spectacle over commercial utility. Digit's 65-pound deadlift is the opposite: unglamorous, practical, and aimed directly at the tasks that actually generate revenue in a warehouse. The two approaches represent fundamentally different philosophies about what humanoid robots are for, and Agility's commercial traction suggests the boring-but-useful strategy is winning right now.
Figure AI has also been aggressively developing its Figure 01 and Figure 02 humanoid platforms for warehouse applications, with BMW as a high-profile deployment partner. Figure has demonstrated object manipulation and task execution in real factory environments, but published details on maximum payload capacity have been limited. Digit's specific 65-pound number gives Agility a concrete, comparable benchmark that Figure will need to match publicly if it wants to compete on specification sheets in front of procurement teams.
The broader sim-to-real transfer methodology Agility used also deserves comparison to work coming out of academic labs. Researchers at Carnegie Mellon and ETH Zurich have published extensively on domain randomization and object-inclusive simulation for legged robots, and Agility's approach fits squarely within that research tradition. What makes the Digit result meaningful is that it happened on a commercial platform actively being deployed, not in a university lab. Bridging that gap between academic technique and commercial execution is exactly where most robotics companies stumble, and Agility appears to have cleared it. For more context on where humanoid robotics sits right now, check the latest AI Agents Daily news coverage on physical AI developments.
FAQ
Q: What is Agility Robotics' Digit robot used for? A: Digit is a humanoid robot designed primarily for warehouse and logistics work. It stands about 5 feet 4 inches tall and is built to move through human-sized spaces, pick up packages, and perform repetitive physical tasks. Agility Robotics sells access to Digit through a service contract model rather than direct hardware sales.
Q: How did Digit learn to deadlift 65 pounds? A: Agility Robotics trained a machine learning control policy using simulation. They included the 65-pound object in the virtual training environment so the model could learn how the weight affects balance, grip forces, and body coordination. That policy was then transferred successfully to the physical Digit robot.
Q: What is the sim-to-real gap in robotics? A: The sim-to-real gap refers to the problem where robots trained in virtual simulations fail to perform correctly when placed in the real world. Simulations are approximations of physics, so subtle differences between the virtual environment and physical reality can cause trained behaviors to break down. Closing this gap is one of the central technical challenges in modern robotics.
Digit's deadlift marks a genuine step forward for commercial humanoid robotics, not because lifting weight is new, but because doing it reliably on a robot already deployed in real warehouses proves the training pipeline works end to end. As Agility Robotics heads into the second half of 2026 with expanding RaaS contracts and a growing capability set, the rest of the industry will be watching closely. Subscribe to the AI Agents Daily weekly newsletter for daily updates on AI agents, tools, and automation.
Get stories like this daily
Free briefing. Curated from 50+ sources. 5-minute read every morning.