Manipulation Is the New Frontier

As humanoid robots master locomotion, the race has shifted toward dexterity, whole-body control, and the science of robotic touch.

May 29, 2026

“Without hands there would be no humanity.” — André Leroi-Gourhan

On May 18, 2026, Boston Dynamics posted a video that, on its surface, was almost mundane. Its electric Atlas humanoid bent its knees, wrapped both arms around a 23-kilogram mini-fridge, lifted it, walked it to a table, and — without moving its feet — rotated its torso a full 180 degrees, to put it down. No backflips. No parkour. A robot moved an appliance across a room.

It is, quietly, one of the most consequential robotics demonstrations in years. Not because lifting a fridge is hard for a machine — a forklift does heavier work — but because of how Atlas did it, and what that “how” tells us about the single hardest unsolved problem in robotics. The industry has a name for it now, and it’s about to dominate the conversation at the world’s largest robotics conference. To understand why, you have to understand the difference between a robot that moves and a robot that works.

What the video actually shows

The instinct is to focus on the strength. That’s the wrong thing to watch. The remarkable part is balance — specifically, what roboticists call whole-body control (or whole-body manipulation).

Think about how a conventional industrial robot handles an object. The arm is the actor; the base is a fixed, immovable platform. The arm reaches, grips, and moves, while everything below stays bolted in place. This works beautifully when the object is light relative to the robot and the environment is perfectly structured. It falls apart the moment the object is heavy, awkward, or unpredictable — because a heavy object hanging off the end of an arm becomes a lever that wants to tip the whole machine over.

A 23-kilogram fridge held at arm’s length is exactly that kind of destabilizing load. To handle it, Atlas can’t treat its legs and torso as a static pedestal. It has to coordinate its legs, hips, torso, and arms as a single integrated system: the knees bend to drop the center of mass, the torso leans to counterweight the load, and the whole body continuously re-solves its own balance in real time as the fridge’s mass shifts. Boston Dynamics’ own framing captured the point bluntly — you cannot lift a fridge just by looking at it and using your hands. The hands are almost the least interesting part. The intelligence is in the coordination of everything else.

Two further details matter. First, the company says Atlas learned the task through large-scale, simulation-based reinforcement learning — running the equivalent of enormous numbers of practice attempts in a virtual world, then transferring the learned policy to real hardware in a matter of weeks. Second, the production Atlas that performed it is a deliberately commercial machine: 56 degrees of freedom, fully rotational joints that let it spin its torso in ways no human spine can, human-scale hands with tactile sensing in the fingers and palms, the ability to swap its own battery, and a rated lift of up to roughly 110 pounds (the company says it handled loaded fridges well beyond the 23 kg shown). This isn’t a research stunt. It’s a capability demo aimed squarely at deployment on Hyundai production lines.

The hardest problem in robotics is the one humans find easy

There’s a famous observation in robotics, often traced to Moravec’s paradox: the things that are hard for humans (chess, calculus, logic) turned out to be relatively easy for machines, while the things any toddler does effortlessly (walking across a cluttered room, picking up a strange object, manipulating it in-hand) turned out to be staggeringly hard.

Manipulation is the purest expression of this paradox. A robot that can do a backflip — something almost no human can do — still needs millions of simulated attempts to learn a lifting task a person masters in a single try. Locomotion, balance, jumping: these have largely fallen to modern methods. But reliable, general-purpose physical interaction with the world — grasping unfamiliar objects, handling deformable or heavy ones, recovering when a grip slips — remains the field’s stubborn frontier.

This is why so many in the field now describe manipulation as the remaining bottleneck for humanoid robots. The machines can already move through the world impressively. Working within it — touching, lifting, assembling, manipulating — is the part that separates a viral demo from a useful coworker. The Atlas fridge video is significant precisely because it’s a credible step on that specific axis, not the well-trodden axis of mobility.

A short history of teaching robots to touch

The push toward dexterity is not new. It’s a decades-long arc, and the fridge demo sits near its leading edge.

The story begins in the analytic era. Early dexterous manipulation research, from the 1980s onward, was grounded in classical control and analytic models — researchers tried to mathematically describe contact forces, friction, and grasp stability, and to compute optimal grips from first principles. The iconic hardware ambitions of this period — multi-fingered robotic hands like the Utah/MIT and Stanford/JPL hands — were extraordinary feats of mechatronics, but they were brittle: they worked when the world matched the model and struggled when it didn’t.

The limitation was fundamental. The real world is too rich in contact dynamics, material variation, and uncertainty to fully pre-compute. A grasp that’s stable in theory slips in practice; an object that’s rigid in the model deforms in reality. Analytic methods gave robots precise control in structured settings and left them helpless in unstructured ones.

The last decade rewrote the playbook with learning. Instead of deriving grasps from physics equations, robots began learning them from data — first through imitation learning (copying demonstrations), then reinforcement learning (practicing through trial and error, increasingly in simulation), and most recently through large vision-language-action (VLA) models that borrow the scaling logic of large language models and apply it to physical action. The research frontier has moved toward in-hand reorientation, dexterous regrasping, and general-purpose hand control learned from sources as unconventional as egocentric human video. ICRA’s 2025 best paper on manipulation was, tellingly, about a unified representation for cross-embodiment dexterous grasping — learning grasps that generalize across different robot hands and different objects, exactly the generalization that analytic methods could never achieve.

The Atlas fridge demo is a product of this learning paradigm — millions of simulated repetitions, a policy that transfers to hardware — fused with the older, hard-won engineering of balance and whole-body control. It represents the two halves of the field’s history converging: classical control’s mastery of dynamics, plus modern learning’s capacity to generalize.

The measurement problem: how do you prove a robot is dexterous?

Here is where the story takes a turn that rarely makes the headlines but matters enormously. A viral video proves a robot can do a task once, on a stage, under conditions the company controlled. It does not prove the robot is reliable, that the result is reproducible, or that one company’s “dexterous” robot is more capable than another’s. The clips are unfalsifiable by design.

This is the gap that benchmarking and standards work exists to close — and it’s quietly become one of the most important enabling efforts in the entire field.

The central player here is the U.S. National Institute of Standards and Technology (NIST), working through ASTM International Committee F45 on Robotics, Automation, and Autonomous Systems — specifically its subcommittee F45.05 on Grasping and Manipulation. The premise of this work is deceptively simple: if the industry wants robotic manipulation to be adopted at scale, buyers and developers need a common, repeatable way to measure performance — not marketing claims, but standardized tests anyone can run and compare.

F45.05’s effort is organized around several working groups, each tackling a different slice of the problem: benchmarking grasp-type end-effector technology, benchmarking assembly capabilities, and benchmarking mobile-manipulator performance. The most tangible output is NIST’s Assembly Task Boards (ATBs) — physical, standardized fixtures studded with representative manufacturing tasks (inserting pegs, seating connectors, manipulating deformable objects like belts and wires) that a robot must complete under a specified protocol, with time and success recorded. First introduced to the community as competitions back in 2017, these task boards turn the vague question “is this robot dexterous?” into a concrete, reproducible score. The work explicitly aims to match robotic hand capabilities to end-user needs and to give developers objective insight into where their hardware and software fall short.

This is the unglamorous infrastructure beneath the flashy demos. It’s the difference between “look what our robot did” and “here is how our robot scores on a test you can run yourself.” NIST has even framed part of its mission as a roadmap to advance measurement science in robot dexterity and manipulation — an acknowledgment that you cannot improve, or trust, what you cannot measure.

ManipulationNet and the authenticity problem

The benchmarking challenge has sharpened as the field has shifted to learning-based systems, because those systems are notoriously good at looking impressive in conditions that don’t generalize. A 2026 benchmarking effort, ManipulationNet, frames the core issue as one of authenticity: a meaningful benchmark must require different systems to perform the same task, with identical objects, under specified protocols, with reported results that faithfully reflect what actually happened.

The deeper point ManipulationNet raises is the tension between simulation and reality. Simulation-based benchmarks are reproducible and scale infinitely — you can run millions of trials — but their imperfect approximations of contact dynamics (friction, deformation, the messy physics of touch) mean strong simulation results don’t necessarily reflect true real-world capability. Real-world evaluation has physical fidelity but is hard to scale and hard to keep consistent across labs. This is the same sim-to-real gap that Boston Dynamics had to cross to get Atlas’s simulated fridge-lifting policy to work on actual hardware: in simulation, floor friction and motor response and sensor readings are perfect; in the real world, they aren’t, and behaviors that look reliable in sim routinely fail on metal. Benchmarking infrastructure like ManipulationNet and the NIST task boards exists precisely to keep the field honest about which capabilities are real.

The £57 million bet on robot bodies

Measurement tells you where the field stands. Funding determines where it goes — and one of the boldest financial bets on this exact problem comes from the UK’s Advanced Research and Invention Agency (ARIA), the government-backed research funder loosely modeled on America’s DARPA but without the defense remit, created to back high-risk, high-reward science that’s too speculative or interdisciplinary to find money elsewhere.

In 2024, ARIA launched a £57 million Robot Dexterity programme, led by Programme Director Jenny Read, a physicist and computational visual neuroscientist, and a professor of vision science at Newcastle University. What makes the programme so telling is what it chose not to fund. When ARIA made its most ambitious robotics investment, it deliberately skipped over AI software almost entirely and aimed the money at something less fashionable and, in its view, more neglected: robot bodies — the hands, actuators, sensors, and control that let a machine physically touch the world.

The reasoning is the central thesis of this entire story, stated as a funding strategy. As ARIA frames it, breakthroughs in AI are transforming what robots could do, but compared with the staggering advances in computation, the development of robot bodies has stalled for decades — and today’s robots still can’t match the flexibility, speed, and precision of human manipulation, which renders them useless for many of the difficult or dangerous jobs where they’re most needed. The programme explicitly ties the urgency to demographics: the share of the world’s population aged 65 and over is projected to triple by 2100, even as labour shortages in physically demanding work deepen. Read has put it plainly — we need robots to help face this century’s challenges, most notably an aging population. The bet, in other words, is that AI will be ready before robot hands are, and that hardware is now the binding constraint.

The money is spread deliberately wide and weird. ARIA funded 25 research teams across 26 projects, spanning startups, universities, public research institutions, and major companies — and notably funds teams worldwide, not just in the UK, where it believes the work can benefit the country. The portfolio is a tour of the frontier this article has described: Artimus Robotics won £1.5M to push artificial-muscle (soft actuation) technology into dexterous manipulators with the University of Bristol; the University of Bath received £2.2M to model soft robotic manipulators, reframing hardware design as a quasi-evolutionary process that co-optimizes a robot’s body and its control software together; and Cambridge runs three projects spanning hardware and advanced simulation. The first solicitation asked, bluntly, for manipulators demonstrating dexterity that far exceeds what’s possible today — substantial leaps in performance and robustness without introducing deal-breakers in cost, size, or scalability — alongside new techniques in actuation and haptic sensing.

This is the same instinct as NIST’s benchmarking work, pointed at a different lever. NIST builds the ruler to measure dexterity; ARIA writes nine-figure checks to advance the thing being measured — and both proceed from the identical conviction that the body, not the brain, is now the bottleneck. It’s also no coincidence that Read chose the Humanoids Summit in London as the venue to discuss the programme’s global push: the funding bet and the research community are converging on the same stage.

Why Vienna, next week, matters

All of this comes to a head at ICRA 2026 — the IEEE International Conference on Robotics and Automation, the world’s foremost robotics conference, running June 1–5 in Vienna, Austria, and drawing on the order of 7,000 of the field’s top researchers and companies.

Manipulation and dexterity are not a side track at this event; they are arguably the central theme. The momentum is visible across the field’s recent calendar: dexterous manipulation dominated discussion at IEEE Humanoids, the UK Grasping and Manipulation community has been convening specifically around these problems, and ICRA’s own programming reflects the shift — dedicated workshops on active perception for manipulation (the idea that a robot should physically interact with objects to gather the information it needs to manipulate them), sessions on tactile sensing, and a steady stream of papers on dexterous grasping and whole-body control. The 2025 best-paper award going to cross-embodiment dexterous grasping was a leading indicator; the 2026 program is the confirmation.

So the Atlas fridge video lands at a precise moment. It’s a vivid, public, industrial demonstration of whole-body manipulation arriving in the same window that the research community gathers in Vienna to push exactly this frontier — and that the standards community, through NIST and ASTM F45.05, builds the measurement scaffolding to tell which advances are real. The demo, the research, and the benchmarks are three faces of one story.

Getting Our “Hands” Around It

It’s tempting to watch a humanoid carry a refrigerator and file it under “robots are getting strong.” That misses the point entirely. Strength was never the hard part. The hard part is the thing the video makes look easy: a machine coordinating its entire body to manage an unpredictable physical load, balancing in real time, and doing it from a policy it taught itself in simulation.

That capability — full-body, dexterous, reliable physical interaction with the world — is the bottleneck standing between today’s impressive humanoids and genuinely useful ones. Closing it requires several things converging right now: the engineering breakthroughs (Atlas), the research community pushing generalizable dexterity (ICRA, Humanoids, the UK workshops), the deliberate funding bets on robot bodies over software (ARIA’s £57M Robot Dexterity programme), and the unglamorous measurement science that proves which claims hold up (NIST, ASTM F45.05, ManipulationNet).

The robots have learned to move. Learning to work — to touch the world as deftly as we do — is the decade’s real project. A humanoid lifting a fridge, then turning to set it down, is what the first real progress on that project looks like.

Robot News Of The Week

3D-printable humanoid legs let robotics experiments run wild

Hugging Face is betting that the future of humanoid robotics may not begin with a $100,000 machine, but with a $2,500 set of open-source robot legs. The company’s new LeRobot Humanoid project combines 3D-printable parts, off-the-shelf hardware, and simulation-ready software into a platform designed less for flashy demos and more for experimentation, learning, and rapid iteration. The release reflects a growing push toward affordable, accessible robotics development at a moment when commercial humanoid costs remain high and investor hype is colliding with economic reality. The bigger question: could open-source hardware do for robotics what Linux once did for software?

Hyundai Motor Company Plans Massive Deployment of Humanoid Robots Across U.S. Factories

Hyundai Motor Company plans to deploy more than 25,000 Atlas humanoid robots across its manufacturing operations beginning in 2028. Developed by Boston Dynamics, the robots will perform physically demanding factory tasks as Hyundai pushes to transform itself from automaker into a Physical AI and robotics company.

Robot Research Of The Week

It looks like a sea urchin, but this strange 20-legged machine is rewriting what robots can do

For decades, robotics has borrowed its shapes from nature — dogs, insects, even humans. But researchers at Duke University are asking a different question: what if the future’s most capable robots don’t look familiar at all? Their answer is Argus, a strange sea-urchin-like machine with 20 telescoping legs and no true front or back. Built around a new concept called “dynamic isotropy,” the robot can move, stabilize, and adapt almost equally well in every direction. The project hints at a broader shift in robotics away from biomimicry and toward mathematically optimized bodies designed not to imitate life, but to outperform traditional assumptions about movement itself.

Clinical trial seeks to advance intuitive assistive robotics for people with paralysis

Rice University and Baylor College of Medicine are joining the BrainGate consortium to advance a future where people with paralysis can control robotic assistive devices directly with their thoughts. The clinical trial will focus on decoding neural signals to help individuals with tetraplegia perform everyday tasks like eating and drinking independently using robotic systems. The effort reflects the growing convergence of brain-computer interfaces, AI, and robotics — not just to restore movement, but to restore autonomy. As neuroprosthetics rapidly evolve from lab experiments into clinical technologies, the project signals how assistive robotics may become one of the most transformative real-world applications of physical AI.

Robot Workforce Story Of The Week

VCU to offer new undergraduate degree in robotics and autonomous systems engineering

Virginia Commonwealth University is betting that the future engineer won’t just write code or design machines — they’ll build intelligent systems that move through the real world. This fall, VCU will launch Virginia’s first undergraduate degree in Robotics and Autonomous Systems Engineering, a hands-on, interdisciplinary program built around AI, robotics, embedded systems, and real-world deployment. With project-based learning, a dedicated robotics lab, and faculty spanning multiple engineering disciplines, the program aims to prepare students for a rapidly expanding robotics economy expected to reshape industries from manufacturing and healthcare to defense and transportation.