1X’s NEO Robot Hands Highlight the Race for Dependable Home Automation

MIT Technology Review has shifted its AI Hype Index toward an awkward conflict: robots now look capable of cooking dinner, despite unresolved questions about autonomy.

The immediate catalyst is 1X, the Norwegian-American robotics company developing the NEO home humanoid. Its new robotic hands can manipulate food, tools, clothing, game controllers, and other objects that defeat conventional grippers. The company presents those abilities as foundations for useful household work.

Yet the impressive fingers do not settle the larger argument. A hand can execute a difficult motion while the robot still depends on human direction. The real contest is therefore not humans against dexterous machines. It is the polished promise of household autonomy against the unglamorous engineering required to deliver it reliably.

That distinction matters as anxiety about AI employment grows. More than 200 economists, researchers, and technology leaders recently warned that increasingly capable AI systems could cause large-scale job displacement. A robot that appears ready to handle dinner adds a physical dimension to a debate previously dominated by office software.

However, a controlled demonstration is not the same as an unsupervised evening in a crowded kitchen. Homes contain reflective surfaces, unfamiliar packages, children, pets, spills, and objects placed in unexpected locations. A useful household robot must handle that disorder without turning every exception into a support call.

The latest AI Hype Index captures this tension. The most revealing AI story is no longer a single spectacular behavior. It is whether companies can transform those behaviors into dependable services that people trust around food, possessions, and family members.

MIT Technology Review Spots the Unsexy Work Behind Robot Hands

1X has made the robotic hand look less like a laboratory attachment and more like an interface for everyday life.

The company describes its latest NEO hands as having 25 actuated degrees of freedom. A degree of freedom is an independently controlled direction of movement, such as bending one finger joint. More controllable movement gives the hand additional ways to shape itself around irregular objects.

Of those 25 degrees, 22 sit within the fingers and palm, while three belong to the wrist. The design uses tendons to transmit force from motors into the fingers. It resembles the basic logic of a human hand, although the materials, sensing, and control systems differ.

1X says the hands use relatively low gear ratios between roughly 5:1 and 15:1. Lower gearing can let a joint respond more naturally when an external force pushes against it. Engineers call this backdrivability, meaning contact can move the mechanism instead of meeting a rigid wall.

The company’s detailed hand specifications also describe tactile sensing across contact surfaces. Tactile sensors estimate information such as pressure and slippage when the robot touches an object. That feedback helps the control system adjust its grip before a grape, glass, or plastic package escapes.

1X reports peak torque of 3.5 newton-meters at one thumb joint and 2.6 newton-meters at major finger joints. It also claims distal flexion forces reaching 45 newtons. These figures describe mechanical capability, not successful completion rates across household tasks.

The hands are also described as waterproof to the IP68 standard and made with food-safe materials. Those attributes matter because household work involves water, grease, crumbs, and repeated cleaning. A robot that cannot wash its hands would have limited value around a kitchen.

The company’s videos place these specifications into recognizable situations. The hands pick individual grapes, manipulate small screws, unzip clothing, use a game controller, and open snack packaging. Such actions demand different combinations of force, alignment, fingertip control, and feedback.

These are not merely decorative flourishes. Opening flexible packaging is harder than lifting a rigid box because the object changes shape during contact. Picking a grape requires enough force to hold it, but not enough to crush it.

Still, the demonstrations establish a component’s range rather than a complete robot’s independence. They show that the hand can perform motions under the conditions used for recording. They do not reveal how often NEO can identify, plan, and complete each task without human help.

That is the first important reversal in the hype cycle. The flashy visual directs attention toward the fingers. The decisive questions sit behind them, inside perception, task planning, recovery behavior, and remote support.

Dexterity Changes What Household Automation Can Attempt

A capable hand expands the robot’s task list, but useful automation depends on the entire chain between an instruction and a safe result.

Most home robots operate within narrow physical boundaries. Robot vacuums move across floors, while smart appliances control one machine or process. They succeed because manufacturers reduce the number of objects and decisions each device must manage.

A humanoid takes the opposite approach. It enters spaces designed for human bodies and tries to use handles, switches, drawers, utensils, and packages already present. That strategy avoids rebuilding the home, but it gives the robot a far larger collection of edge cases.

Hands are central because homes contain tools made for human fingers. A parallel gripper can move some boxes and containers. It struggles with tied bags, stacked plates, fabric, small fasteners, slippery food, and controls requiring several fingers.

NEO’s expanded dexterity gives 1X a chance to address those categories without specialized attachments. A single hand design can theoretically handle dishes, laundry, groceries, doors, and entertainment devices. That flexibility is the commercial appeal behind general-purpose humanoids.

Cooking is an especially useful test because it combines nearly every difficulty. The robot must locate ingredients, interpret their condition, select tools, manage contamination, regulate force, and respond when an object moves unexpectedly. Heat and sharp implements make mistakes more consequential.

Even a basic meal contains many hidden decisions. A person notices when packaging tears badly, a vegetable rolls away, or a plate sits too close to an edge. Humans correct those problems without treating each one as a new engineering project.

A robot needs a perception system that recognizes the changed state. It then needs a policy, meaning a learned or programmed method for choosing the next action. Finally, its hand must execute that action with appropriate speed and force.

Failure recovery separates an interesting demo from a practical appliance. If NEO drops a utensil, the robot must recognize the failure and decide whether the utensil remains usable. It may need to clean the item, choose another one, or ask for assistance.

This sequence explains why dexterity represents progress without proving autonomy. Better hands increase the number of actions available to the system. They do not guarantee that its software will choose the correct action at the correct moment.

The distinction also clarifies why “unsexy AI” is an effective description. The visible achievement is a robot opening a package. The underlying work includes sensor calibration, motor control, object tracking, safety limits, data collection, and thousands of uneventful repetitions.

Those repetitions are difficult to market. A video of a robot completing the same task safely across varied kitchens lacks the surprise of a first demonstration. Yet repetition produces the evidence buyers need before trusting a machine in private spaces.

Developers should therefore read the hand announcement as a change in the available hardware interface. It creates a wider foundation for manipulation models and training data. It does not remove the need for system-level evaluation.

Enterprise buyers already apply this distinction to software agents. A model completing one workflow once is interesting. A system completing it consistently, documenting exceptions, and escalating uncertainty is useful.

Household robotics now faces the same standard, with higher physical stakes. A mistaken document can be corrected. A mistaken movement near boiling water, medication, or a sleeping pet can cause immediate harm.

The Real Opponent Is the Demo-to-Autonomy Gap

The central contest is not 1X against another humanoid maker. It is 1X’s consumer promise against the limited evidence available today.

NEO is positioned as a home robot rather than a factory machine. 1X began taking consumer orders in 2025 and says initial United States deliveries will start in 2026. That timeline puts pressure on the company to convert demonstrations into repeatable household performance.

The challenge is unusually difficult because factory deployments simplify the environment. Industrial robots often work with fixed stations, known objects, safety barriers, or closely managed processes. Engineers can measure a constrained task and redesign the workspace around failure points.

A household offers fewer controls. Lighting changes throughout the day. Furniture moves, products arrive in different packaging, and people interrupt tasks. Every customer effectively supplies a new operating environment.

That variation increases the value of a generalist model. 1X calls its system Redwood AI, which it says learns and repeats tasks. A generalist model aims to transfer experience across multiple activities instead of receiving separate programming for each motion.

However, generalization must be measured rather than inferred from varied demonstrations. A robot performing ten recorded tasks might use one broadly capable policy. It might also depend on significant task-specific preparation, human guidance, or carefully arranged objects.

The company has openly included teleoperation in its development model. Teleoperation lets a human operator control or guide the robot remotely when its autonomous system cannot complete a task. The resulting interaction can also create training data for future autonomous behavior.

That approach is technically reasonable. Humans supply corrections in situations where the machine lacks experience. Over time, the company can use those examples to improve perception and action policies.

It also creates the largest unresolved issue around NEO. If a person must intervene frequently, the product behaves less like autonomous labor and more like a physical interface for remote labor. The robot’s usefulness then depends on operator availability, latency, privacy controls, and service economics.

A hands-on assessment noted that 1X ultimately seeks full automation, while acknowledging the role of teleoperation in current development. That gap should not be treated as proof of deception. It should be treated as the main performance metric.

The most useful disclosure would be an intervention rate. This rate measures how often a human must assist the robot during normal use. It can be reported per task, operating hour, or successful completion.

A declining intervention rate would show that data collection and model updates are producing greater independence. A stable rate would suggest that demonstrations are expanding faster than autonomy. Without that measurement, viewers must interpret selectively edited videos with caution.

Completion rates need equal attention. A robot that finishes a chore without intervention might still take too long, damage objects, or leave the workspace unsafe. The relevant benchmark combines success, time, safety, and recovery quality.

Competitive videos offer limited help because companies choose different tasks and conditions. Tesla has promoted Optimus through manipulation and movement demonstrations. Figure has emphasized workplace deployments and its broader vision for general-purpose humanoids.

Boston Dynamics, Apptronik, Sanctuary AI, and several Chinese manufacturers are also pursuing capable humanoid systems. Their designs vary in target market, actuation, hand complexity, training methods, and business strategy. A dramatic clip cannot normalize those differences.

1X’s consumer-first approach makes the autonomy gap particularly visible. A factory customer can supervise early systems and redesign a process around them. A household expects a product to work without employing a robotics team.

This is why the primary opponent is promise versus reality, not 1X versus Tesla. Another company’s impressive demonstration does not validate NEO. Nor does a competitor’s delay prove that 1X has solved domestic autonomy.

The decisive evidence must come from ordinary use. NEO needs to perform in customer homes, with unfamiliar layouts and unplanned interruptions. Until that evidence arrives, the hands are an important mechanical achievement attached to an unsettled service proposition.

AI Job Anxiety Has Entered the Kitchen

Dexterous robots make automation anxiety more tangible, but the labor impact depends on deployment speed and task economics.

The timing of 1X’s demonstrations intensifies their cultural impact. In July 2026, more than 200 economists, AI researchers, and technology leaders signed a four-sentence letter about AI’s economic consequences. Sixteen Nobel Prize winners were among the signatories.

The economists’ warning says AI may become radically more capable within the next decade. It describes both large-scale job displacement and potential improvements in living standards. The signatories call for institutions that direct AI toward complementing people.

That wording is more cautious than many social media summaries suggest. The letter does not assert that a specific number of jobs will disappear. It warns that rapid capability gains could transform the economy faster than institutions can adapt.

Generative AI initially made this fear feel concentrated in language work. Writers, programmers, analysts, designers, and customer-support teams saw software perform parts of their workflows. Physical occupations appeared to have a different barrier because machines still struggled with irregular environments.

Dexterous humanoids challenge that comforting division. A robot able to handle flexible objects, tools, and food appears relevant to hospitality, cleaning, care work, retail, warehousing, and light manufacturing. The visual connection between capability and work becomes immediate.

Appearance can still outrun economic reality. A robot must complete tasks reliably enough to justify installation, supervision, maintenance, insurance, and training. A human performing several related duties may remain more adaptable than a machine optimized around selected demonstrations.

The likely near-term unit of automation is therefore the task, not the occupation. A humanoid might move laundry, retrieve objects, or clear surfaces while a person handles judgment-heavy work. That pattern can complement workers, reduce staffing needs, or do both.

The result depends on how organizations redesign jobs. If companies use robots to remove dangerous or exhausting tasks, workers can spend more time on communication and exceptions. If they use the same systems primarily to reduce head count, the distributional outcome changes.

Home robots introduce another dynamic. Unpaid domestic work does not appear in employment statistics like a restaurant job does. Yet reducing household labor can affect caregivers, personal aides, cleaning services, and how families divide their time.

For older adults or people with mobility limitations, a dependable robot could support greater independence. Retrieving items, moving laundry, and handling simple cleaning might reduce reliance on relatives. These benefits require unusually strong safety and accessibility testing.

A remotely assisted robot complicates the labor story further. Automation in one home may create operator work somewhere else. That arrangement can distribute expertise, but it can also conceal human labor behind a machine’s appearance.

Transparency matters because customers and policymakers need to know whether a system acts autonomously. They also need to know when a person can access cameras, microphones, or controls inside a private residence. Consent must cover the actual service, not its futuristic presentation.

The employment debate therefore cannot be settled by a hand demonstration. Mechanical capability establishes that a wider range of tasks is technically reachable. Deployment data determines whether those tasks become economically automatable.

MIT Technology Review’s framing works because it resists both simple conclusions. NEO is not merely hype, since its hardware addresses serious manipulation problems. It is not evidence of mass replacement, since reliable autonomy and broad deployment remain unproven.

What the Robot Hand Videos Do Not Measure

The strongest case for skepticism concerns missing operational evidence, not whether the fingers moved in a recorded demonstration.

A product video compresses time and removes failed attempts. Viewers rarely see how many trials preceded the final sequence. They also cannot determine whether a human controlled the robot, selected intermediate actions, or reset objects outside the frame.

This does not make demonstrations worthless. They reveal physical range, design priorities, and tasks a company considers achievable. They are weak evidence for reliability because the audience lacks the denominator behind each successful attempt.

For household robots, that denominator should include hundreds or thousands of task attempts across varied environments. Companies should report independent completion, human-assisted completion, failure, and safety-stop rates. They should also describe how test homes differ from training locations.

Manipulation benchmarks need realistic variation. Packaging should change in stiffness, size, reflectivity, and opening mechanism. Dishes should arrive wet, stacked, misplaced, or partly obscured.

Food handling raises additional questions. Food-safe materials help, but they do not guarantee hygienic behavior. The system must avoid cross-contamination, recognize dirty surfaces, and know when cleaning has failed.

Water resistance similarly addresses one component risk. It does not establish that the complete robot can operate safely near sinks, appliances, and people. A household certification program will need to evaluate the entire system.

Privacy is another hidden variable. Teleoperation requires sensory access sufficient for an operator to understand the robot’s surroundings. In a home, that access can expose faces, conversations, documents, medication, financial information, and daily routines.

1X says users control when expert assistance occurs. The practical safeguards still need examination through shipping software, customer agreements, retention policies, and security testing. Marketing assurances cannot substitute for technical and contractual controls.

Cybersecurity becomes physical security when a connected robot can move objects. Authentication failures, software vulnerabilities, or compromised operator accounts could produce consequences beyond data loss. Companies must design limited privileges, visible indicators, logs, and rapid shutdown behavior.

Mechanical safety also depends on more than peak force. Engineers need predictable force limits during faults, collisions, sensor errors, and network failures. A compliant hand is useful, but the arms and full body must respond safely.

Maintenance will test the design’s durability. Tendons experience wear, tactile skins encounter cuts, and household contaminants enter small gaps. The relevant question is not whether a new hand performs well, but how performance changes after sustained use.

Support burden offers another reality check. Early customers will discover unusual combinations of objects and tasks. If every novel situation requires an operator or engineer, deployment may generate valuable data while remaining expensive to support.

This possibility does not invalidate the strategy. Real homes can produce training examples that laboratories cannot anticipate. Early deployment may be the mechanism through which NEO becomes more independent.

It does mean that customers could function as participants in a long development process. Companies should make that status explicit. Buyers need to understand which tasks work independently, which require scheduled assistance, and which remain unsupported.

Independent reviews will provide better evidence than launch videos. Reviewers should test the same chore repeatedly, change the environment, interrupt the robot, and observe recovery. They should also document when remote help enters the process.

The industry needs benchmarks that resemble appliance testing more than model leaderboards. Household utility depends on uptime, consistency, cleaning, serviceability, noise, energy use, and safe failure. Intelligence scores alone will not capture those qualities.

This is the real meaning of unsexy AI. Progress will appear as fewer interventions, fewer damaged objects, longer maintenance intervals, and clearer privacy controls. None of those metrics produces an instantly viral clip.

Yet those metrics determine whether NEO becomes useful infrastructure or an expensive demonstration platform. They also determine whether humanoids create broad economic pressure or remain limited to enthusiasts and structured trials.

Three Signals Will Reveal Whether Unsexy AI Works

The next phase should be judged through customer evidence, autonomy measurements, and competitive verification.

The first signal is delivery into ordinary United States homes. 1X says deliveries begin in 2026, making customer deployment the nearest test of its claims. Shipping dates matter less than what recipients can consistently accomplish.

Early owners should document task duration, independent completion, remote assistance, and failure recovery. One successful meal preparation will attract attention. Repeated results across different kitchens will establish whether the capability transfers beyond company-controlled settings.

Evidence from accessibility use cases deserves particular attention. A robot retrieving objects or handling laundry for someone with limited mobility creates measurable utility. It also subjects safety, voice control, and reliability to demanding real-world conditions.

If customers complete useful chores with modest setup and limited intervention, the promise-versus-reality gap will narrow. If most complex tasks require expert assistance, NEO will remain a supervised platform. Both outcomes would provide valuable information unavailable from demonstrations.

The second signal is transparent autonomy reporting. 1X should publish intervention rates, task success, recovery rates, and the conditions used for evaluation. Ideally, those numbers should distinguish familiar tasks from new objects and environments.

The direction of change matters more than a single headline figure. A steady decline in interventions would support the company’s learning strategy. It would show that remote demonstrations and household experience are feeding a system that becomes more independent.

A flat intervention rate would weaken that interpretation. It would suggest that the robot can perform more chore types while still leaning heavily on people. That outcome might support a remote service, but it would not satisfy the usual meaning of household automation.

The third signal is comparable evidence from competitors. Tesla, Figure, Apptronik, and other developers will continue releasing humanoid demonstrations. The useful response would be measured household or workplace performance under disclosed conditions.

Competitive pressure can improve both hardware and reporting. If one company publishes credible completion and intervention data, others will face pressure to match it. The race could move from cinematic demonstrations toward operational evidence.

A shared benchmark would strengthen that shift. It could include unpacking groceries, sorting laundry, clearing a table, cleaning a spill, and responding to interruptions. Each task should use varied objects and impose strict safety limits.

The benchmark should also identify teleoperation and human preparation. Those details are not embarrassing exceptions. They are necessary variables for understanding how the system works.

MIT Technology Review’s AI Hype Index is most useful when it keeps these distinctions visible. The index should not reduce every demonstration to empty marketing. It should also resist treating a component breakthrough as proof of a finished autonomous worker.

For developers, the lesson is to measure the complete workflow. Better models and hands matter when they reduce intervention and recover from uncertainty. For enterprise buyers, the lesson is to demand operational rates rather than highlight reels.

Knowledge workers should also watch this transition. Robotics exposes the same weakness found in software agents: completing an idealized task is easier than owning a messy process. Reliability, permissions, escalation, and auditability remain the difficult work.

The next viral NEO video will probably showcase another recognizable human skill. Before sharing it as evidence of widespread automation, ask three questions. Did the robot act independently, how often did it succeed, and what happened when the environment changed?

Those answers will determine whether “unsexy AI” becomes dependable household infrastructure. They will also show whether MIT Technology Review is tracking a genuine transition or another impressive stop on the hype cycle.