What Is AI Humanoid Robots Explained How They Learn Tasks?

Introduction

A growing number of readers now ask one question when they meet these machines: what is AI humanoid robots explained how they learn tasks? The short answer is that a humanoid robot learns skills from human demonstrations, simulated practice, and large pretrained models instead of hand-written rules. That shift is attracting serious capital, and Goldman Sachs Research projects a humanoid robot market of $38 billion by 2035 in its latest forecast. Behind that number sits a quieter technical story about how a robot goes from watching a person fold a shirt to folding one itself. This guide walks through imitation learning, reinforcement learning, simulation, and the vision-language-action models that now tie those methods together. It also covers the safety standards, privacy questions, and economic trade-offs that decide whether learning humanoids reach factories and homes. By the end, you will know what these robots can do, how they get there, and where the claims outrun the evidence.

Quick Answers on AI Humanoid Robots and How They Learn Tasks

What is AI humanoid robots explained how they learn tasks?

It explains how human-shaped robots acquire skills through imitation of demonstrations, reinforcement learning in simulation, and vision-language-action models, so they can handle new objects and instructions without hand-written code for every motion.

How do humanoid robots learn a new task?

Most learn from teleoperated or filmed human demonstrations, then refine the skill with reinforcement learning in simulation. Pretrained vision-language models add scene understanding, so one policy can follow many spoken instructions.

Do humanoid robots learn on their own?

Only partly. Today’s humanoids still depend on human demonstrations, curated data, and remote operators. Autonomous improvement happens mostly inside simulators, and real deployments need supervision and tightly bounded tasks.

Key Takeaways

• Humanoid robots learn mainly through imitation of human demonstrations, reinforcement learning in simulation, and large pretrained vision-language-action models.
• Teleoperation data is the scarce ingredient, and Figure’s Helix model needed only about 500 hours of it to control a whole upper body.
• Simulation is cheap and fast but suffers from a sim-to-real gap, so real-world data and randomized training conditions stay essential.
• Safety standards, privacy, and uneven reliability, not raw intelligence, now decide how quickly learning humanoids spread.

What Is a Humanoid Robot That Learns Tasks With AI?

What is AI humanoid robots explained how they learn tasks? It describes human-shaped machines that acquire skills from demonstrations, simulation, and trial and error, not hand-written rules. A neural network policy converts camera and joint signals into motion.

The Anatomy of a Learning Humanoid: Sensors, Actuators, and Compute

Every learning humanoid is built from three layers that the learning software has to coordinate. The first layer is perception, which usually combines head-mounted cameras, wrist cameras, an inertial measurement unit that tracks tilt, and encoders that report every joint angle. Some platforms add force and torque sensors in the wrists or fingertip tactile pads, which give a policy a sense of touch that cameras alone cannot supply. The second layer is the body itself, where electric motors, gearboxes, and linkages turn commands into movement across dozens of joints. Boston Dynamics describes its latest Atlas as having 50 degrees of freedom in its public engineering update. Each of those joints is one more number the learned policy must predict many times per second. Every added joint multiplies the amount of data a robot needs before its behavior looks smooth and reliable.

The third layer is compute, and it determines how large a learned model the robot can carry on its own frame. Early humanoids ran hand-tuned controllers on modest embedded processors, while current systems carry GPU-class chips so that a neural network can run on board without a network connection. Figure reports that its Helix model runs on dual low-power embedded GPUs inside the robot, which removes cloud latency from the control loop. That design matters because a robot that falls over waits for no server, and balance corrections must arrive within a few milliseconds. Most teams therefore split the work, with a slow reasoning model deciding what to do and a fast control network deciding exactly how the joints should move.

The hands deserve separate attention, since manipulation is where most learned skills succeed or fail. A hand is an example of an end effector in robotics, and designers trade finger count, strength, speed, and cost against each other. A five-fingered hand with many motors offers human-like flexibility but is expensive, fragile, and harder to learn control for. A simpler gripper learns faster and breaks less often, yet it cannot rotate a screw or thread a cable the way a human hand does. These hardware choices shape what a learning algorithm can even attempt, which is why software and body are always designed together.

From Hand-Coded Rules to Learned Behavior: A Short History

Building on that hardware picture, it helps to see how control software for walking and grasping evolved. The first generation of humanoids, built from the 1970s through the 2000s, relied on carefully derived equations of motion and hand-tuned gains. Engineers wrote down where the robot’s center of mass should travel, and controllers chased that trajectory with painstaking precision. These robots walked impressively on flat floors, but any surprise, such as a rug, a push, or a slightly different floor height, could topple them. The core limitation of rule-based control is that every situation must be anticipated by a human engineer in advance.

Machine learning entered gradually, first in perception and later in motion. Computer vision networks replaced hand-written image filters, letting robots recognize cups, handles, and doors from raw pixels, a progression explored in this overview of computer vision in robotics. Reinforcement learning then showed that locomotion policies could be discovered in simulation instead of derived by hand, and demonstrations showed that arms could copy human motions directly from data. The most recent step merged these threads, using large pretrained language and vision models to give robots a general understanding of objects and instructions. Each stage moved more of the robot’s competence out of human-written equations and into data, which is exactly why data collection has become the central bottleneck in the field today. Anyone who types “what is AI humanoid robots explained how they learn tasks?” into a search engine is really asking how that handover from engineers to data happened.

Imitation Learning: Copying Human Demonstrations

Stepping back from the history, imitation learning is the most direct answer to how a robot learns a task. A person performs the task while sensors record what the robot would see and which joint commands would reproduce the motion. The learning algorithm then trains a neural network to predict those commands from the observations, a technique usually called behavior cloning. If the demonstrations are good and plentiful, the resulting policy can reproduce the task without any explicit programming of the steps. Imitation learning works because it converts a hard control problem into a supervised learning problem that engineers already know how to solve.

The weakness of pure behavior cloning is compounding error, because small mistakes push the robot into states that no demonstration ever covered. Once the robot drifts outside its training distribution, its next prediction is less reliable, and the drift grows until the task fails. Researchers fight this with several tricks, including collecting recovery demonstrations, adding noise during data collection, and predicting short sequences of future actions instead of one step at a time. Boston Dynamics reported that its early Atlas policies could not recover from a dropped part until the team demonstrated recoveries and retrained. That finding appears in the company’s large behavior model update. Generative policy designs, such as diffusion transformers, also help by representing many valid ways of doing the same job rather than averaging them into a hesitant blur.

Data volume and quality matter as much as the algorithm. A demonstration set filled with hesitations, regrasps, and inconsistent strategies teaches a robot to hesitate, regrasp, and behave inconsistently. Teams therefore invest in operator training, consistent task definitions, and careful labeling, since cleaner demonstrations beat larger noisy ones in many experiments. The best pipelines also close the loop by evaluating the trained policy and finding exactly where it fails. They then collect more demonstrations aimed specifically at those failures, which is far cheaper than blindly gathering more of everything.

Teleoperation and Motion Capture: Teaching a Robot by Doing

Turning to the data source, teleoperation is how most high-quality humanoid demonstrations are gathered today. A human operator wears a virtual reality headset, hand controllers, and sometimes trackers on the feet and chest, and the robot mirrors every movement in real time. The robot’s own cameras and joint sensors record what happens, which means the dataset matches the robot’s body exactly, with no translation from human anatomy needed. Boston Dynamics describes VR headsets, controllers, and foot and chest trackers feeding its Atlas demonstrations, with the robot’s stance and stepping mirroring the operator. This approach produces precisely the observation-and-action pairs that behavior cloning needs.

Motion capture offers a second route, where humans move freely while cameras or suits record their poses. Software then retargets the human motion to the robot’s proportions, a step that is harder than it sounds because arm lengths, joint limits, and balance differ. The CoRL 2025 system called TWIST, the Teleoperated Whole-Body Imitation System, shows one approach, combining retargeted motion capture with a controller trained by reinforcement learning and behavior cloning. The authors report that privileged future motion frames and real-world capture data improved tracking, and that one unified network produced whole-body manipulation, locomotion, and expressive movement. Whole-body teleoperation matters because tasks like lifting a box require legs, torso, and arms to cooperate.

The cost of this method is human time, and that cost shapes every business plan built on teleoperation. Each hour of demonstration needs a trained operator, a working robot, and a person to reset the scene between attempts. Scaling to many thousands of hours is therefore slow and expensive, even for well-funded companies. Operators also tire, and fatigue shows up as slower, shakier motion that the policy then imitates faithfully. Latency and the lack of touch feedback make delicate tasks hard, since a headset controller cannot tell the operator how firmly a cup is being squeezed. For these reasons, companies are actively searching for cheaper sources of demonstration, including human video and synthetic data, which later sections cover.

Quality control becomes a discipline of its own once a fleet of operators is involved. Teams write task scripts, define success criteria, and review recordings so that a policy is never trained on a failed or inconsistent attempt. Some labs also tag each episode with the operator, the robot, and the lighting conditions, which lets engineers discover that a policy performs poorly only in one setting. Standard task definitions let data from different sites be pooled without confusing the model. Careful teams also keep a held-out set of demonstrations that is never used for training, so they can measure real progress instead of memorization. For builders, the demonstration pipeline is the practical heart of any honest answer to “what is AI humanoid robots explained how they learn tasks?” in real projects.

Reinforcement Learning: Trial, Error, and Reward

Shifting from copying to discovering, reinforcement learning lets a robot find its own solution by maximizing a reward. The designer specifies what counts as good, such as staying upright, moving at a commanded speed, and using little energy. The robot then tries millions of variations, strengthening whichever behaviors earn more reward and abandoning those that do not. For a general explanation of the family of methods, this overview of supervised, unsupervised, and reinforcement learning is a useful companion. Reinforcement learning is the preferred tool for locomotion because walking is easy to score but nearly impossible to demonstrate precisely. A human cannot tell a robot how much torque each ankle should apply on every step, yet a reward for staying upright captures the goal in a single line.

The practical difficulty is that real robots cannot afford millions of falls. A single tumble can bend a leg or shatter a sensor, and every attempt takes real time, so nearly all humanoid reinforcement learning happens in simulation. Berkeley researchers trained a transformer policy for humanoid walking using large-scale reinforcement learning on randomized environments in IsaacGym, collecting about 10 billion samples in roughly a day. The resulting policy transferred zero-shot to real outdoor terrain, including sidewalks, running tracks, and grass. The robot even developed a human-like arm swing that nobody programmed, and it recovered from a foot trapped on an unseen step. That result illustrates the power of simulation, because no physical robot could have lived through ten billion steps.

Reward design is the quiet art behind these results, and it is also where projects most often go wrong. A poorly designed reward invites exploitation, where the robot finds a loophole, such as vibrating forward instead of walking, that scores well but looks absurd. Engineers therefore combine many reward terms for balance, smoothness, foot clearance, and energy use, then tune their weights through repeated experiments. They also add penalties that discourage jerky actuator commands, which protects motors and makes the simulated gait easier to transfer. Reinforcement learning is weaker for manipulation tasks with sparse success signals. A robot that must open a drawer, find a mug, and place it on a shelf rarely stumbles onto that sequence by chance. Practitioners solve this by starting from demonstrations and then fine-tuning with reinforcement learning, which combines the direction of imitation with the polish of trial and error.

Simulation and the Sim-to-Real Gap

Looking at where all that trial and error happens, simulation is the training ground that makes modern humanoid learning affordable. A physics engine models gravity, contact, friction, and motor dynamics, so thousands of virtual robots can practice in parallel on a single GPU cluster. Time can run faster than reality, failures cost nothing, and engineers can reset a scene instantly. NVIDIA describes a workflow in which Isaac Lab trains policies with reinforcement and imitation learning, then Isaac Sim validates them in physically accurate environments before deployment. The same company says a synthetic motion generation pipeline can produce large datasets from a small number of human demonstrations. That extends the reach of expensive real recordings and lets a small team test far more ideas per week.

The catch is the sim-to-real gap, the set of differences between the simulator and the physical world. Real motors have backlash and delay, real floors have unmodeled friction, and real cameras suffer glare and motion blur that a clean render lacks. A policy tuned to the simulator’s quirks can fail the moment it meets those differences. Closing the sim-to-real gap is largely a matter of making the simulator uncertain enough that the policy never relies on its exact details. The main technique is domain randomization, which varies mass, friction, motor strength, sensor noise, and terrain every episode. A policy that survives thousands of slightly wrong worlds learns to be robust to the real one, which is simply another slightly wrong world.

Other tools include system identification, which measures the real robot to tune the simulator, and history-conditioned networks that infer hidden properties from recent motion. The Berkeley team hypothesized that its transformer used observation and action history to adapt at test time without updating weights, a behavior they likened to in-context learning. Another approach to realism is world models, such as those described in NVIDIA’s Cosmos work on humanoid navigation, which generate photorealistic scenes for training perception. Even with these tools, contact-rich manipulation remains hard to simulate, because soft cloth, liquids, cables, and squishy food behave in ways physics engines approximate poorly. Engineers often split the problem, training walking and balance almost entirely in simulation while collecting real data for hand tasks. This division of labor explains why a humanoid can walk gracefully across rubble yet still struggle to tie a shoelace. The sim-to-real gap is therefore a central part of the answer to “what is AI humanoid robots explained how they learn tasks?” in practice.

Vision-Language-Action Models: The New Brain Stack

Building on those training methods, the most important architectural idea of the last two years is the vision-language-action model. A vision-language-action model, often shortened to VLA, takes camera images and a written or spoken instruction as input and produces robot actions as output. It starts from a vision-language model that was pretrained on internet-scale text and images, so it already knows what a mug, a drawer, and the word “tidy” mean. Robot data is then added on top, which teaches the network how those concepts translate into finger and arm motion. The result is a single network that can look at a cluttered counter, read a request, and decide which joints to move. A VLA lets a robot inherit common-sense knowledge from the internet instead of learning every object and phrase from robot demonstrations alone.

Figure’s Helix model shows the design in a concrete form. According to the company’s Helix announcement, a 7-billion-parameter open-weight vision-language model it calls System 2 runs at roughly 7 to 9 Hz and handles scene understanding and language. An 80-million-parameter transformer called System 1 turns that model’s latent output into continuous motor commands at 200 Hz. The system controls 35 degrees of freedom, covering fingers, wrists, torso, and head, and it was trained on about 500 hours of teleoperated data. Figure says that amount is under 5 percent of earlier VLA datasets, and the whole model runs on dual low-power embedded GPUs inside the robot. Two robots running identical weights also coordinated on a shared task without being assigned separate roles.

The dual-frequency split is the key engineering insight, and it resembles the way people separate deliberate planning from reflexive movement. A large model is too slow to steer joints directly, because balance and contact demand corrections hundreds of times per second. A small fast network therefore handles the reflexes while a big slow network supplies intent. That idea fits the broader trend described in this explainer on the rise of multimodal AI. Many policies also predict short chunks of future actions at once, then execute part of the chunk before planning again. Boston Dynamics, for example, describes a policy of about 450 million parameters that predicts 48 action steps at 30 Hz and executes 24 before the next inference. Chunking smooths motion and hides small delays in the larger model, which is why it appears in so many recent designs.

Impressive demonstrations should not be confused with measured reliability in real working conditions. Figure’s announcement reports picking up thousands of novel household items from simple prompts, but it publishes no success rates, benchmark scores, or baseline comparisons. Language grounding can still fail when an instruction is ambiguous or an object looks unfamiliar, and a model can confidently grasp the wrong item. Latency from the larger model can also delay reactions when a scene changes quickly, such as when a person walks into the workspace. Independent evaluations on shared benchmarks are therefore the next milestone before anyone should treat VLA performance as settled. Until then, the sensible reading is that VLAs are a promising architecture whose real-world robustness is still being measured.

Learning From Human Video and Synthetic Data

Looking for cheaper teachers than teleoperators, several companies now try to learn from ordinary human video. Tesla is reported to be moving Optimus training away from motion-capture suits and teleoperation toward a camera-first approach. eWeek’s account of the shift says workers wear helmets and backpacks carrying five cameras while they fold shirts and pick up objects. The footage then trains the robot to imitate those actions, mirroring the vision-only method Tesla used for its driving software. The appeal is scale, since a person can wear a camera rig all day without needing a robot, a headset, or a reset crew. Video is cheap to collect but carries no direct record of the forces and joint commands that produced the motion.

That missing information is the central research problem for video-based learning. A video shows where a hand went, but it does not show how hard the fingers squeezed or which torques held the arm steady. Software must therefore infer actions from pixels and retarget them to a robot whose body proportions differ from a human’s. Critics quoted in the same eWeek report argue that teleoperation lets a robot learn through physical interaction, which video alone cannot provide. Supporters counter that a multi-angle rig may capture fine joint and finger positions with more precision than a cruder tracker. The honest conclusion is that video helps most for pretraining and for learning what a task looks like. A smaller amount of robot-specific data then teaches the body how to execute it.

Synthetic data fills a different gap by generating scenes and motions that nobody had to record. NVIDIA’s robotics training models are one example of tools that expand a handful of demonstrations into many varied training episodes. Generated data can vary lighting, object placement, and camera angle, which makes a policy less brittle when it meets a real kitchen or factory aisle. NVIDIA also describes a synthetic motion generation pipeline that can produce large datasets from a small number of human demonstrations. The risk is that artifacts in generated scenes can teach a model false regularities, such as a texture that never appears in reality. Teams therefore validate every synthetic gain on a physical robot before trusting it, and they keep a core of real data as the anchor.

Foundation Models and Cross-Embodiment Transfer

Stepping back from single robots, researchers have asked whether one model can learn from many different bodies at once. The Open X-Embodiment collaboration pooled data from 22 robot embodiments, drawing on 60 existing datasets and more than one million real robot trajectories. The authors report that RT-1-X outperformed the original methods trained on individual datasets by 50 percent in small-data domains. They also report that a 55-billion-parameter RT-2-X model beat RT-2 by a factor of three on emergent skill evaluations. These results suggest positive transfer, meaning that data from other robots can help a given robot even though their hands and arms differ. The finding matters for humanoids, because every new body would otherwise need its own expensive dataset from scratch.

NVIDIA’s GR00T N1 applies the same philosophy to humanoids specifically. The GR00T N1 paper describes an open foundation model with a vision-language module for reasoning and a diffusion transformer module for real-time motor actions, trained jointly end to end. Its training mix combines real-robot trajectories, human videos, and synthetically generated datasets, so no single data source carries the whole burden. The model was deployed on the Fourier GR-1 humanoid for language-conditioned bimanual manipulation, where the authors report strong performance and high data efficiency. A foundation model turns robot learning from a per-robot, per-task project into a fine-tuning problem on top of shared pretrained knowledge. Instead of starting from zero, a new robot team can adapt a pretrained model using a modest set of its own demonstrations.

Transfer has real limits that the impressive headline numbers can easily hide. Robots differ in camera placement, joint ranges, control rates, and gripper design, so actions learned on one body must be translated, and that translation loses detail. Concepts from transfer learning in machine learning apply directly, including the risk of negative transfer when the source and target tasks are too far apart. Most successful systems so far use cross-embodiment data for pretraining and then specialize with data from the target robot. Open datasets also vary widely in quality, which makes careful curation as important as raw quantity. Researchers therefore treat dataset documentation, filtering, and balanced sampling as first-class engineering work rather than housekeeping.

Whole-Body Control: Balance, Walking, and Manipulation

Turning to the body itself, whole-body control is the layer that keeps a humanoid upright while its hands do useful work. A two-legged robot is an inverted pendulum with many joints, and every arm movement shifts its center of mass and threatens its balance. Classical controllers solved this with model-based optimization, and many commercial robots still rely on those methods for their lowest-level stability. Learned controllers increasingly supplement or replace them, because a neural network can handle messy contact and uneven ground without an explicit model of each case. The TWIST work mentioned earlier trains one unified network that handles locomotion, legged manipulation, and expressive motion together. The hardest skills for a humanoid are not the flashy ones but the ones that demand balance and manipulation at the same time.

Athletic tasks show how far learning has come in coordinating a whole body. A system described in this report on a robot that learned badminton in simulation had to coordinate footwork, arm swing, and timing against a moving shuttle. The same coordination appears when a humanoid carries a box while stepping around an obstacle or opens a door while keeping its balance. In each case a single learned policy must reason about the whole body rather than one limb at a time. Researchers often layer a fast balance controller beneath a slower task policy, so that a mistake in the upper level does not immediately topple the robot. That layering is also where safety limits are easiest to enforce, since the lowest layer can refuse any command that would break stability.

Putting Humanoid Learning Into Practice: An Implementation Roadmap

Moving on to deployment, a team that wants a humanoid to learn a task follows a repeatable loop. Boston Dynamics describes the working loop as four distinct steps. Teams collect teleoperated demonstrations on hardware and in simulation, process and annotate the data, train one neural network on all tasks, and evaluate on a test suite. The evaluation results then guide what data to collect next, so the loop tightens over time. Success depends on defining the task precisely before any data is recorded, including what counts as completion and which failures are acceptable. Teams that skip this step end up with inconsistent demonstrations that no algorithm can reconcile. The most common reason learning projects stall is not model quality but unclear task definitions and inconsistent data collection.

The first practical decision is whether to fine-tune an existing foundation model or to train from scratch. Fine-tuning a pretrained model usually needs far less data, and it preserves broad knowledge about objects and language that a small dataset could never teach. The approach relies on the same ideas as transfer learning in machine learning. Training from scratch makes sense only when a robot’s body or sensors differ so much that no existing model applies. Teams should plan the evaluation suite in parallel with data collection, including edge cases such as odd lighting, cluttered tables, and unexpected human movement. A good suite measures not just success rate but also time per attempt, number of human interventions, and recovery from failure. These numbers let a manager compare a new policy against the last one and decide whether it is safe to release.

Deployment adds operational concerns that research papers rarely mention or measure. Robots need maintenance schedules, spare parts, safe charging areas, and trained staff who can pause or reset them without risk. Many customers prefer the robotics as a service model, in which a vendor retains ownership and handles upkeep for a recurring fee. Agility Robotics used this approach with GXO, deploying its Digit humanoid under a robots-as-a-service agreement that the companies described as the first of its kind. Whatever the commercial model, teams should start with one narrow, high-volume task, measure it honestly, and expand only after it works reliably. Treating the first deployment as a learning exercise for the organization, not just for the robot, prevents costly surprises later. A team that follows this roadmap can give a concrete, measured reply to “what is AI humanoid robots explained how they learn tasks?” before spending heavily on hardware.

Where Learned Humanoid Skills Fall Short: What the Risks Are

Despite the progress, learned humanoid skills remain brittle in ways that polished demonstrations conceal. A policy trained on thousands of successful attempts has seen few examples of things going wrong. A surprise such as a slippery handle or a moved shelf can therefore cause failure. Boston Dynamics acknowledged that its early policies could not recover from disturbances until the team collected recovery demonstrations and retrained. The company also notes that collecting data beyond stationary manipulation, while keeping motion smooth and responsive, remains difficult. Task dynamics sometimes prevent speeding up execution, which limits how quickly a learned robot can match a human worker. A learned skill is only as general as the data it came from, and robot data is still scarce compared with the text and images that trained chatbots.

Reliability evidence is still thin across the entire field of learning humanoids. Figure’s Helix announcement presents demonstrations rather than success rates or benchmark comparisons, and independent reproduction of such claims is rare. Morgan Stanley’s China shipment analysis was accompanied by a note, in the AI Weekly summary, that public reporting still lacks hard data on reliability and uptime. Hardware wears out as well, and Figure said the forearm was the main failure point of its robots during an 11-month factory pilot. A model can be flawless in software and still lose a day of work to a broken cable or an overheated actuator. Buyers should therefore ask vendors for measured uptime, intervention rates, and mean time between failures before signing a contract.

Generalization failures can also be silent, which makes them dangerous. A robot may complete a task correctly in the lab and then misread a label, a reflective surface, or a transparent container on a customer’s floor. Because learned policies are opaque neural networks, engineers cannot inspect a written rule to see why a particular decision was made. They must instead probe behavior with structured tests, log every failure, and retrain, which is slow and expensive. Monitoring in production is just as important as testing before release, because the real world keeps producing situations that no test suite anticipated. Teams that treat deployment as a continuing evaluation exercise catch these problems earlier.

Security is a related concern that grows as robots follow natural language. A policy that accepts spoken or typed instructions can be steered by a malicious or mistaken request, and adversarial inputs can confuse its perception. Teams mitigate this with permission layers, restricted task menus, and hard-coded safety limits beneath the learned policy. Responsible deployment treats the neural network as one component inside a larger system with independent guardrails, not as the final authority over a powerful machine. Logging every instruction and action also creates an audit trail that investigators can review after an incident.

Safety, Standards, and Working Beside People

Given the limits above, safety engineering becomes the gating factor for putting learned robots next to people. A humanoid is dynamically stable, which means it must actively balance to stay upright, and cutting its power makes it fall rather than simply stop. Falling humanoids weighing 70 kilograms or more can seriously injure workers, according to an explainer on ISO 25785-1. Existing industrial robot standards assume a machine that does not tip over, so stability as a hazard fell between older rules. The emerging ISO 25785-1 draft exists to close that gap by covering industrial mobile robots whose stability depends on active control. It applies to humanoids, quadrupeds, and balancing wheeled platforms that operate on ground surfaces in industrial settings.

The draft addresses fall detection, controlled descent, testing under rated load, payload effects on balance, and recovery after instability. As of early 2026 it was a working draft that first appeared in May 2025, with ratification industry-estimated at 18 to 36 months after the draft. That timeline means large-scale deployment of collaborative humanoids is unlikely before 2027, according to the same explainer. Until a standard is final, employers rely on general duties, risk assessments, and documented procedures for fall zones and energy control. A broader primer on robot safety standards explains how these rules fit with established machine guarding practice. Facilities should calculate and mark fall zones, write energy control procedures, train operators, and log incidents.

Learning adds a safety question that fixed-program robots never raised. If a policy changes after retraining, the validated behavior of last month may differ from today’s, so each update needs its own regression tests. Developers can wrap learned policies inside hard limits on force, speed, and workspace, so that even a wrong action cannot exceed a safe envelope. Google DeepMind’s Gemini Robotics team describes a similar layering, in which a model can judge whether an action is safe in context while separate low-level controllers handle collision avoidance. Practical coexistence also depends on behavior that people can predict, such as robots that signal their intentions, slow down near humans, and stop reliably when asked. Those habits earn public trust faster than any technical specification or marketing claim.

Ethics, Privacy, and the Human Labor Behind the Demos

Stepping back from engineering, the ethics of learning humanoids begin with an uncomfortable fact about how many of them learn. Teleoperation data comes from people, and in some products remote operators do the work while the robot supposedly gets smarter. The 1X NEO home robot, for example, uses human operators wearing Quest 3 headsets during scheduled “Expert Mode” intervals, according to Road to VR’s report. Teleoperation gives the robot’s network real-world experience for chores such as loading a dishwasher, and the company says quality will improve as data accumulates. The same report notes that during a press demonstration the robot performed no tasks autonomously. Anyone buying a learning home robot is also buying a window into their home for the people and systems that train it.

Privacy therefore becomes a design variable rather than a footnote. The report says owners can restrict the robot from certain rooms and may be able to blur people from its view. The company’s chief executive describes privacy as a sliding scale in which sharing more data makes the robot more useful. That framing puts the burden on households to trade intimacy for competence. Consent becomes complicated when guests, children, or roommates appear in footage they never agreed to share. Questions about where recordings are stored, who can view them, and how long they are kept deserve written answers before purchase. The broader debate about humanoid robots in the home raises these questions long before the products are widespread.

Labor ethics matter as well, and they extend beyond the household. Operators who record demonstrations, label failures, and supervise robots perform real work, often under short-term contracts and with little public visibility. Fair compensation, clear data ownership, and transparent disclosure when a human is steering the robot are reasonable expectations. Companies should say plainly when a demonstration is autonomous and when it is teleoperated, since blurred claims erode public trust. Bias and accountability complete the picture, because a robot trained mostly in a few kitchens may handle other households poorly. Responsibility for errors must be assigned clearly among manufacturer, operator, and owner, and regulators have only begun to consider those questions. A complete reply to “what is AI humanoid robots explained how they learn tasks?” must therefore include who supplied the data and who watched the camera feeds.

Jobs, Wages, and Economic Impact

Looking at the labor-market consequences, learning humanoids are aimed first at work that is repetitive, physically demanding, or hazardous. Goldman Sachs Research points to structured industrial settings such as electric vehicle assembly and component sorting, and to dangerous, dirty, and dull work including mining and chemical manufacturing. It estimates that automating 5 to 15 percent of car manufacturing and dangerous jobs could create demand for 1.1 to 3.5 million humanoid units. Those figures are scenarios rather than forecasts of displacement, and the same analysis stresses that a viable mass-produced general-purpose humanoid is not yet proven. The nearer-term effect is more likely a shift in tasks within jobs than the wholesale replacement of workers. Roles that combine physical dexterity with judgment, such as maintenance and field service, are the hardest for a learned policy to match.

History offers a cautious guide to what may come next in this labor transition. Earlier waves of industrial automation eliminated some roles while creating others in maintenance, programming, and supervision, though the gains were uneven across regions and skill levels. The related discussion of how robots affect jobs covers those patterns in more detail. Wages may rise for workers who supervise fleets and fall for tasks a robot can replicate cheaply. Policy choices on retraining, wage insurance, and transition support will matter as much as the technology itself. Employers that involve workers early in pilots tend to find safer workflows and fewer surprises.

Cost, Business Models, and the Market Outlook

Beyond the labor question, the economics determine how quickly any of this scales. Goldman Sachs reports that manufacturing cost per humanoid unit fell from a range of $50,000 to $250,000 to a range of $30,000 to $150,000 in one year. That amounts to a drop of about 40 percent in only twelve months. Morgan Stanley raised its China shipment forecast for 2026 to 50,000 units, according to AI Weekly’s summary, and projected 446,000 annual shipments by 2030. The same report notes that Chinese firms shipped over 80 percent of global humanoids in 2025, a lead explored in coverage of China’s most advanced humanoid robot. Falling hardware cost shifts the bottleneck from building bodies to collecting the data that makes them useful. A cheaper robot only creates value if it can perform enough tasks reliably to justify its purchase price and upkeep.

Business models are still being tested across the industry by vendors and customers. Selling robots outright puts the burden of maintenance and integration on the buyer, while subscription and service models spread cost and align incentives with uptime. Component supply constraints, such as limited capacity for high-precision grinding machines, still slow scaling, according to Goldman. Forecasts also differ widely between analysts, so planners should treat any single number as a scenario and watch for evidence such as repeat orders and measured uptime. A buyer who evaluates total cost of ownership, including training time and downtime, will make a better decision than one who compares sticker prices.

The Future of Humanoid Robots That Learn on the Job

Looking ahead, the most likely path is a gradual expansion from narrow tasks to broader ones as data and models improve. Near-term deployments will resemble the pilots already running, with robots moving totes, loading parts, and carrying bins under human supervision. Google DeepMind reports that Gemini Robotics more than doubles the performance of other leading vision-language-action models on a generalization benchmark. It also says the model can be specialized for humanoids such as Apptronik’s Apollo, as described in the Gemini Robotics announcement. The same announcement describes folding origami and packing a snack into a bag, tasks that demand fine dexterity. The trajectory points toward robots that adapt to new tasks from a few demonstrations and a spoken instruction.

Several technical frontiers will decide how fast that broader adoption actually happens. Tactile sensing and force control would let robots handle fragile and slippery items more reliably. Boston Dynamics lists tactile gripper force control and fast dynamic manipulation among its future work. Better world models could let robots rehearse tasks mentally before acting, which would reduce the amount of costly real-world data required. Continual learning would allow a deployed robot to improve from its own experience without forgetting earlier skills, a capability that remains largely unsolved. Standards such as ISO 25785-1 will shape how and where learning robots may operate near people. Progress on any one of these frontiers would widen the set of tasks that are practical to automate.

Investment signals suggest that capital will keep flowing to the field. Startups focused on general robot intelligence have attracted large funding rounds, as the Physical Intelligence funding round shows, on the thesis that software foundations matter more than any single body. Competition among bodies, models, and data pipelines will likely consolidate around a few platforms, as happened in earlier computing eras. Whether those platforms are open or closed will influence who benefits and how much choice buyers retain. Open models such as GR00T N1 lower the barrier for smaller teams, while closed systems may move faster inside one company. Both approaches will probably coexist for several years as the field matures.

Public acceptance will be the final gate for all of this. People tolerate machines that are useful, predictable, and honest about their limits, and they reject those that surveil, injure, or overpromise. The companies that earn trust will publish reliability data, disclose teleoperation, and design for privacy from the start. If they do, learning humanoids could become ordinary coworkers within a decade, handling the dull and dangerous tasks that people prefer to avoid. If they do not, the technology may stall despite its technical promise, as earlier waves of robotics hype have done. That is the fuller answer to “what is AI humanoid robots explained how they learn tasks?” as it stands today, and it will keep changing as the evidence improves.

Key Insights

• Goldman Sachs Research expects a $38 billion humanoid market by 2035, with 1.4 million units shipped that year, showing investors are betting on robots that learn many tasks.
• Cost per humanoid unit fell about 40 percent in a year, to $30,000 to $150,000, according to Goldman Sachs, so data is now the main constraint.
• Morgan Stanley now forecasts 50,000 humanoid shipments from China in 2026, up from 28,000, and notes that Chinese firms shipped over 80 percent of global units in 2025.
• Figure trained Helix on about 500 hours of teleoperated data, under 5 percent of earlier VLA datasets, showing that a small curated set can drive an entire upper body.
• Pooling more than one million trajectories from 22 robot types let RT-1-X beat single-dataset baselines by 50 percent in small-data domains, evidence that cross-robot data transfers between bodies.
• Berkeley researchers collected about 10 billion simulated samples in roughly a day to train a walking policy that worked zero-shot on real sidewalks, grass, and running tracks.
• Figure’s robots loaded more than 90,000 sheet-metal parts at BMW with accuracy above 99 percent over 11 months, yet the forearm stayed the main hardware failure point.
• The 1X NEO home robot costs $20,000 outright or $500 per month and relies on remote human operators in Expert Mode, so full autonomy is not yet the product.

Taken together, these figures describe an industry whose hardware is getting cheaper faster than its intelligence is getting reliable. Money and shipments are growing quickly, yet the evidence for dependable autonomy still rests on a handful of pilots and demonstrations. Small, carefully curated demonstration sets can control a whole upper body, which suggests that data quality may matter more than sheer volume. Simulation supplies the scale that reinforcement learning needs, while pooled cross-robot datasets show that skills can transfer between different bodies. At the same time, remote operators and hardware wear remind us that today’s deployments still depend on people in the loop. The most useful way to read these numbers is as a map of where learning works well and where human help still carries the load.

The table below compares the four main learning approaches across eight dimensions that matter to teams choosing a method.

Dimension	Imitation learning	Reinforcement learning	Sim-to-real training	VLA foundation models
Main data source	Human demonstrations from teleoperation, motion capture, or video	A reward signal and millions of simulated attempts	Randomized physics simulations plus a little real-world tuning	Internet-scale text and images plus robot demonstrations
Best suited for	Manipulation, such as grasping, placing, and folding	Locomotion, balance, and whole-body motion	Skills too risky or slow to practice on hardware	Following varied spoken instructions across many objects
Human effort required	High, because operators must record clean demonstrations	Moderate, mostly in designing the reward and the simulator	Moderate, mostly in building and randomizing the simulator	Moderate to high, for curating mixed data and fine-tuning
Typical weakness	Compounding errors outside the demonstrated states	Reward loopholes and poor results on sparse-success tasks	Gaps between the simulator and real contact physics	Slow large models and limited public reliability evidence
Compute demand	Low to moderate	High, with billions of simulated samples	High, with many parallel virtual robots	Very high for pretraining, moderate for fine-tuning
Safety during training	Operator is in control and the robot is supervised	Unsafe on hardware, so training stays in simulation	Safe, because failures happen only in software	Depends on the data source and on fine-tuning on hardware
Generalization to new tasks	Limited to the range of the demonstrations	Limited to the trained reward and environment	Improves with broader randomization	Strongest, because of pretrained language and vision knowledge
Typical example	Atlas learning from VR teleoperation	Berkeley’s humanoid walking policy	NVIDIA Isaac Lab workflows	Helix, GR00T N1, and Gemini Robotics

Real-World Practice: Three Deployments of Learning Humanoids

Beyond the market forecasts, three deployments show how the learning methods above behave outside a laboratory. Each one combines demonstrations, simulation, or large pretrained models in a different mix, and each reports numbers that can be checked against a public source. They also come with limits that the original announcements tend to understate. Reading each deployment for what it measured, and what it did not, is the fastest way to separate working systems from polished demos. The examples are drawn from a factory floor, a research robotics program, and a university lab. Together they cover imitation, large behavior models, and reinforcement learning in simulation.

Figure 02 Loading Sheet Metal at BMW

Figure deployed its Figure 02 humanoids at BMW’s Spartanburg plant in South Carolina, where they ran 10-hour shifts loading sheet-metal parts onto welding fixtures. Over 11 months the robots handled more than 90,000 parts and contributed to more than 30,000 BMW X3 vehicles, according to company-reported figures. Placement accuracy stayed above 99 percent with a tolerance of 5 millimeters, and each cycle took 84 seconds including a 37-second load. The robots logged more than 1,250 hours of runtime and walked an estimated 200 miles inside the facility. The forearm was the main failure point, because packing three degrees of freedom, thermal management, and cabling into a human-sized limb proved difficult. The pilot was also limited to lifting parts from bins, since conventional robotic arms still did the welding, and the figures come from the company itself.

Atlas and the Large Behavior Model

Boston Dynamics and Toyota Research Institute trained a language-conditioned large behavior model, a diffusion transformer of roughly 450 million parameters, to control the full Atlas humanoid. The team ran a loop of teleoperated demonstrations, data annotation, joint training on all tasks, and evaluation, using VR headsets and foot and chest trackers for collection. One policy completed a long three-part workshop task, and others handled rope tying, barstool flipping, and a 22-pound tire. Policies also ran 1.5 to 2 times faster without retraining, an increase of 50 to 100 percent in speed, as the Boston Dynamics write-up explains. The approach has limits, because early policies could not recover from a dropped part until the team retrained with recovery demonstrations. The authors also say collecting smooth data beyond stationary manipulation is still difficult, and they do not report total demonstration hours.

Berkeley’s Simulation-Trained Walking Policy

Researchers at Berkeley built a causal transformer that predicts a humanoid’s next action from its history of observations and actions. They trained it with model-free reinforcement learning in IsaacGym, collecting about 10 billion samples in roughly one day, or about 24 hours, across thousands of parallel environments. The policy transferred zero-shot to the real robot and walked on plazas, sidewalks, running tracks, and grass, as shown on the project page. It developed a human-like arm swing and recovered from a foot trapped on an unseen step, without those behaviors being specified. The robot stayed stable when pushed with a stick and when hit by a thrown yoga ball. The work has limits, because the robot is blind, relying only on proprioception, and the training terrains were planes and smooth slopes without discrete obstacles.

Lessons From Three Case Studies in Robot Skill Learning

Moving on to longer stories, three case studies show how data strategy, business model, and commercial deployment shape what learning humanoids can actually do. Each company faced the same underlying problem, which is that robot data is scarce, slow to collect, and expensive to label. Each chose a different solution, ranging from head-mounted cameras to remote human operators to a service contract. Their results are partial, and their critics raise fair questions about evidence. A case study is most useful when it records both what a team built and what remains unproven.

Case Study: Tesla’s Shift to Video Training

Tesla faced a problem common to every humanoid developer, which is that teleoperation and motion-capture data are slow and costly to collect at scale. The company needed to feed Optimus far more examples of ordinary human work than a few operators in suits could produce. In response, it reportedly introduced a vision-only approach, with workers wearing helmets and backpacks carrying five cameras while they fold shirts and pick up objects. The change followed the June 2025 departure of the Optimus director, and Ashok Elluswamy, Tesla’s AI director, has since taken charge, according to eWeek’s reporting. The shift mirrors the camera-first strategy Tesla used for its self-driving software, and the goal is faster data collection. Chief executive Elon Musk said on a January earnings call that Optimus training needs are ultimately at least ten times those of the car.

The impact is not yet measurable, and the case is better read as a bet than a result. Musk claimed in July 2025 that Optimus could generate $30 trillion in revenue, a projection that remains a claim rather than an outcome. Researchers quoted by eWeek raised concerns that video alone cannot teach physical interaction, because teleoperation lets a robot learn through contact and force. Another academic argued that Optimus needs generalized skills instead of memorized actions, since learning each task separately would take too long. Supporters note that multi-angle footage may capture precise finger and joint positions. Whether video-based learning can carry Optimus from simple tasks to real factory work remains unresolved, and the controversy illustrates the trade-off between data scale and data richness.

Case Study: 1X NEO and Teleoperated Learning

1X faced a data problem that most home robot makers share, because no large dataset of household chores exists and homes vary enormously. Its solution was to ship a humanoid called NEO that human operators can control remotely, so that every chore also becomes training data. According to Road to VR, operators wearing Quest 3 headsets take over during scheduled Expert Mode sessions, and owners can list harder tasks they want handled. The company opened reservations for a $200 refundable fee and expects a 2026 US price of $20,000, or $500 per month by subscription. NEO weighs 66 pounds and offers about four hours of battery life, with quick charging. Chief executive Bernt Børnich says the robot’s neural network gains real-world experience from this teleoperation and that quality will improve as more data is collected.

The impact so far is a business model rather than a proven autonomy gain, since no published figure shows what percent of chores NEO completes without a human. The arrangement raises privacy concerns, because a remote person may see inside a home, and the company itself frames privacy as a sliding scale. During a Wall Street Journal demonstration the robot reportedly performed no tasks autonomously, and it cannot lift overly heavy, hot, or sharp items. The reviewer also argued that teleoperation is a stopgap, since headset controllers lack the touch feedback needed for delicate handling and add latency. The case shows the trade-off between getting useful data early and asking customers to accept human eyes in their homes. It also shows why disclosure of teleoperation matters for informed consent.

Case Study: Agility Digit at GXO

Warehouses face a persistent labor problem, because the last meter of material handling, where totes move from mobile robots to conveyors, still depends on people doing repetitive lifting. GXO Logistics needed a way to test whether humanoids could fill that gap without rebuilding its facilities. Agility Robotics and GXO signed a multi-year agreement on June 27, 2024, and launched Digit at the SPANX fulfillment facility. A proof-of-concept pilot in late 2023 preceded the deal, according to Agility’s announcement. The deal was described as the first formal commercial humanoid deployment and the first robots-as-a-service arrangement for humanoids. Digit moves totes from autonomous mobile robots onto conveyors, orchestrated by Agility’s cloud platform, which also integrates with existing automation. The robots were designed to work safely in human spaces, although the announcement gives no certifications.

Agility reported on November 20, 2025, that Digit had moved 100,000 totes at the Flowery Branch, Georgia site, a milestone covered by Interesting Engineering. The figure is a volume count rather than a percent gain or an hours-saved estimate. The coverage does not state how many robots were involved or how long it took. Uptime, error rates, tote weight, and failure modes were also not disclosed. The program is part of a wider GXO test that includes humanoids from Apptronik and Reflex. Critics would reasonably point out that the claim of being first to report measurable volume is the vendor’s own and has not been independently verified. The case therefore shows genuine commercial progress alongside a continuing lack of public reliability data.

Frequently Asked Questions About AI Humanoid Robots and Task Learning