Humanoid Robots: From ASIMO to Tesla Optimus - How Close Are We?

Vídeo em Inglês, publicado pelo canal Daily AI Snapshot.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

The video explores the evolution, achievements, and current realities of humanoid robots, contrasting impressive demonstrations in labs with their actual deployment and economic viability in factories, warehouses, and homes.

Key points

• Humanoid robotics has a long history beginning decades before modern artificial intelligence, with early research projects like Waseda University's Wabot-1 and Honda's Asimo demonstrating bipedal walking and human-robot interaction.
• Technical achievements and impressive demonstrations, such as backflips or dancing routines, do not automatically translate into viable commercial products or all-purpose workers.
• Companies like Figure and Agility Robotics have deployed humanoids into industrial environments, with Figure operating in BMW plants and Digit moving containers in logistics warehouses.
• Economic viability depends on factors like cost per correctly completed task, uptime, hardware maintenance, human supervision, and utilization rates rather than just manufacturing capability.
• Transitioning from structured factory environments to unpredictable home environments presents immense challenges regarding manipulation, safety, variety, and cost justification.

Topics

Humanoid robotsEmbodied AIIndustrial automationWarehouse logisticsHardware and manipulationEconomic viabilityConsumer market

Video transcript

Transcript from available captions; it may contain errors.

Welcome to the daily AI snapshot channel. Today we'll explore the evolution of humanoid robots. From the earliest experiments to today's AI powered machines, a humanoid robot has already helped produce tens of thousands of cars. Another has moved more than a 100,000 containers in a working warehouse. Yet, a machine that can reliably fold your laundry, clean your kitchen, and put everything away remains much harder to build than these impressive industrial milestones might suggest. So, what is real? What is still a carefully controlled demonstration? And how far are we from a robot that can actually help at home? The answer begins long before Tesla, and it is more surprising than the latest viral videos suggest. Humanoid robots did not begin with artificial intelligence breakthroughs in the 2020s. In fact, engineers were building human-shaped machines decades before chat GPT existed, before smartphones, and even before the modern personal computer became common. And some of those earlier robots were remarkably capable. So why are they not already doing our laundry? Our story begins in Japan. In 1973, researchers at Wasida University completed Waybot 1. Widely recognized as one of the first full-scale humanoid robots. It could walk on two legs, grasp objects with its hands, and communicate using simple Japanese. By modern standards, it was extremely slow and limited. But at the time it demonstrated something extraordinary. A machine could combine movement, perception, and communication in a roughly human form. A decade later, Wasida developed Waybot 2, a robot designed to read music and play a keyboard instrument. This was not a household assistant. It was a research project demonstrating how sensors, computing, and precise mechanical control might work together. And that distinction matters. Many early humanoids succeeded at the scientific question they were built to answer, even if they never became products. Then came Honda. Honda began researching bipeedal walking in 1986. Its first experimental machine took around 15 seconds to complete a single step. Through years of prototypes, Honda developed increasingly capable walking robots, culminating in Asimo, introduced in 2000. For anyone following technology in the early 2000s, Asimo looked like a glimpse of the future. The small white robot could walk through public spaces, climb stairs, recognize people, and later run, jump, carry objects, and perform carefully designed service tasks. Some versions could interact with visitors, and navigate without being continuously directed by an operator. And yet, Asimo never became the domestic helper that many people imagined. The problem was not that Honda had built a fake robot. Asimo was an authentic engineering achievement very different from making it useful all day every day. Handling unfamiliar objects, recovering from errors, functioning safely around unpredictable people, and performing enough valuable work to justify the cost were much harder problems. Honda eventually ended Osimo's public demonstration era and redirected its robotics expertise into other technologies. The machine did not become a commercial household product, but its research legacy survived. Sony followed a different path with Qrio, a small humanoid that could walk, dance, and demonstrate sophisticated balance for its time. It was impressive as an entertainment and research platform, but it never went on sale. In 2006, Sony ended further development as part of a broader business restructuring. Decades later, Sony still had no plans to bring Qrio back. These stories reveal an uncomfortable reality. Technical brilliance alone does not guarantee a viable business. Another chapter came from Aldaran Robotics, the company behind now and later the widely recognizable Pepper robot associated with SoftBank. Now became a useful platform for education and research. Pepper with its expressive gestures and conversational features appeared in stores, offices, and public-f facing demonstrations. Unlike many research prototypes, these machines actually reached customers. But the gap between greeting visitors and delivering meaningful labor savings proved enormous. Pepper could attract attention and guide people through scripted interactions. It could not replace a versatile employee or manage complex unpredictable physical tasks. According to Reuters, production was halted around 2020 after the robot struggled to build a sufficiently large customer base. Alderan itself closed in 2025. That does not erase the real educational and commercial uses of its robots. It shows how difficult it can be to convert an engaging demonstration into a sustainable market. Meanwhile, researchers were building humanoids for much tougher environments. South Korea's Hubo family and its DRC Hubo successor were designed to navigate spaces intended for human beings. In 2015, a team from the Korea Advanced Institute of Science and Technology won the Darker Robotics Challenge with DRC Hubo, a competition inspired by the challenge of sending robots into dangerous disaster zones. This was an important milestone, but those competition machines also made the limitations visible. Tasks that humans might complete in minutes could require careful planning, slow movements, and specialized engineering. The robots were showing what was possible, not proving they were ready to replace disaster response teams. And Boston Dynamics was already working on Atlas long before the current humanoid boom. Its earlier hydraulic robots performed dramatic mobility demonstrations, including jumps and complex obstacle courses. That work helped push locomotion and balance forward. But a robot capable of a spectacular backflip is not automatically a robot that can reliably sort hundreds of different parts through an 8-hour shift. So, did these older projects fail? The answer depends on what success means. Wobbot helped establish the field. OIMO advanced walking and human robot interaction. Qrio explored compact, agile machines. Mayo became a genuine research and education platform. Pepper demonstrated both the attraction and the limits of social robots. Hubo and Atlas expanded what machines could physically accomplish. What largely failed was the expectation that those achievements would quickly turn into affordable generalpurpose robotic workers. Earlier generations faced expensive motors and sensors, limited batteries, enormous engineering complexity, and software that struggled whenever the world stopped following a predictable script. They could perform extraordinary actions, but often only inside narrow conditions. That brings us to the real question behind today's humanoid excitement. Have the fundamental problems finally been solved? or have faster chips, cheaper components, modern machine learning, and vastly better marketing simply made the next generation look more convincing. To explore that, we need to look beyond the promotional videos and examine what today's machines can actually do. And one of the most closely watched contenders comes from a company better known for electric cars, Tesla. Tesla's Optimus may be the most recognizable humanoid robot in the world, but recognition is not the same as a proven business model. When Tesla introduced its humanoid ambition in 2021, the presentation included a person dancing in a robot costume. It was an unusual beginning for a project that Elon Musk would later describe as potentially more valuable than Tesla's car business. Subsequent years brought actual prototypes, improved walking, delicate hand movements, and demonstrations of object sorting and factory related tasks. Tesla also has an unusual advantage. It knows how to design motors batteries cameras computing systems, and products intended for mass manufacturer. If humanoid robots can be built cheaply at scale, those skills might matter as much as artificial intelligence. But what precisely has Optimus proved? Its demonstrations show meaningful engineering progress. They do not by themselves establish that a fully autonomous robot can finish an entire factory shift at commercially useful speed without intervention. Some public interactions with Optimus have involved human teleoperation. A robot being remotely guided is still a genuine technological achievement. Behavior is not evidence that software has mastered the same task independently. The important unanswered questions are about repeatability, intervention rates, safety, uptime, and cost per completed job. A company may test robots inside its own factories long before there is independently documented evidence of widespread productive deployment. This distinction becomes clearer when we look at Figure, an American robotics company founded in 2022. Figure attracted attention with a humanoid that could respond to spoken requests, reason about objects in front of it, and manipulate items on a table. Over time, the company's emphasis shifted from impressive conversations toward integrated vision, movement, and control. The goal is not merely to tell a robot what to do. It is to make the robot connect images and language with coordinated physical action. Figure has an advantage that is easier to measure than a beautifully edited demonstration. Real work recorded at an outside customer's factory. In 2025, Figure02 operated in BMW's Spartanberg plant in South Carolina, positioning sheet metal components for automotive production. According to figures reported by BMW itself, the project covered roughly 1,250 operating hours and more than 90,000 handled components supporting production of more than 30,000 BMW X3 vehicles. These are meaningful industrial results. They also need careful interpretation. Supporting the production of 30,000 cars does not mean a robot assembled 30,000 cars. It performed a particular repetitive task within a production process built and supervised by humans. Figure02's run ended and the company moved to the next generation. In 2026, BMW and Figure described testing Figure03 on more complicated logistic sequencing tasks, identifying parts that are not always perfectly positioned, picking them up, and arranging them in the correct order. This is promising because factories are full of variation. But an announced demonstration of a new workflow is not automatically proof of months of successful unattended production. We should judge the new robot by the same standard as its predecessor. Customer confirmed output duration, reliability, and operating cost. Figures progress also introduces a subtle but important point. Independent verification exists on a spectrum. BMW confirming a robot's work inside BMW's plant is much stronger evidence than figure publishing a video. Yet even BMW's statements do not amount to an independent standardized benchmark testing every capability figure advertises. Factory customer confirmation is valuable, but it is not the same as universal competence. Now consider a robot that appears rather different from Optimus or Figure. Agility Robotics Digit has bird-like legs and a design optimized for carrying and moving containers. It does not try very hard to look like a person in a way that makes it more interesting. Digit's commercial story has centered on logistics rather than dramatic demonstrations. In 2024, GXO signed a multi-year agreement to use Digit at a warehouse serving the clothing brand Spanx. Digit moved totes within a defined workflow working alongside other warehouse equipment. Agility later reported a milestone of more than 100,000 totes moved at that facility. GXO's publicly acknowledged commercial involvement makes the deployment particularly significant. We are not just watching a prototype being shown to investors. We are looking at an organization trying to incorporate humanoid style machines into daily operations. Standardized containers is a narrower assignment than organizing a cluttered bedroom, preparing dinner, or taking care of an elderly person. It is a job with known objects and repeatable routes. Robots still need charging, maintenance, oversight, and contingency plans when something fails. commercial operation is an achievement, not the end of the engineering story. Then there is Boston Dynamics, arguably the company that did more than any other to convince the public that humanoid robots could move with astonishing athletic ability. For years, its hydraulic atlas ran, jumped, performed back flips, and recovered from difficult movements. Those videos were groundbreaking demonstrations of mechanics, balance, and control. Yet, the research atlas was not a mass-produced worker. In 2024, Boston Dynamics retired the hydraulic version and introduced a completely electric successor designed around industrial applications. In January 2026, the company unveiled a product version of Electric Atlas with deployments planned for Hyundai and Google DeepMind. Later in the year, Boston Dynamics opened a training and application center at Hyundai's manufacturing complex in Georgia. Atlas robots were being trained for industrial workflows, including moving and sequencing automotive parts. This is a concrete step toward commercialization, but a training facility must not be confused with proof of a large, profitable robot workforce already running an automotive assembly line. The Boston Dynamic story also reveals a fascinating paradox. A humanoid with joints that rotate farther than human joints may be more useful precisely because it does not have to copy our biological limitations. The ideal factory worker might have two legs and two arms, but it does not necessarily need to move like us in every detail. China's unitry offers a very different path. Its relatively affordable humanoid platforms, including the G1, have put sophisticated looking bipedal hardware into the hands of more laboratories, developers, and robotics enthusiasts. Unitry has showcased fast movements, martial arts inspired routines, and striking improvements in balance. The availability of commercial developer platforms can accelerate research because more teams can experiment with real machines instead of building every motor and joint from scratch. However, buying a humanoid platform is not the same as buying an autonomous employee. A base robot's advertised price may exclude essential computing, customization support programming and the labor required to make it useful. A dramatic dance routine proves specific motion skills, not competence at loading a dishwasher. Some spectacular footage across the industry uses pre-recorded choreography or carefully prepared conditions. Whether a robot can cope with a task it did not rehearse in a room it did not memorize with objects that are slightly different from the ones used during training. Other companies matter as well. Aptronics Apollo has been developed with an emphasis on industrial usefulness and partnerships with large manufacturers and logistics companies. Sanctuary AI has concentrated on dextrous manipulation and ways to teach robots physical skills. China's UVtech has pursued humanoids for industrial settings, while companies such as Forier and several newer entrance are experimenting with different combinations of humanlike bodies, teleaoperation, and learned control. Their announcements help show how broad the field has become. But partnerships, pilot plans, and promotional demonstrations should never be placed in the same category as independently observed sustained production. There is another twist in this competition. Some machines, called humanoids, move on wheels. Others have legs shaped quite differently from ours. This is not necessarily a compromise. A warehouse floor is flat and predictable. Wheels can be simpler, cheaper, and more energyefficient than walking. Legs make sense when a robot must use stairs, step over obstacles, or work in spaces that cannot be redesigned. If wheels are good enough for the job, insisting on humanlike walking may be an expensive aesthetic preference. To understand why this competition has accelerated, we need to talk about embodied artificial intelligence. A language model learns statistical relationships in text. An embodied system must connect perception to action in the physical world. It sees a mug, estimates where it is, decides how to approach it, moves an arm, adjusts its grip, and reacts when the mug starts to slip. Some systems use vision, language, and action models to turn instructions and camera images into movements. Others combine those models with traditional motion planning, and control. Robots can learn from human demonstrations teleoperation simulation, reinforcement learning, and their own attempts. The challenge is that the real world does not politely wait for a model to finish thinking. Objects move, lighting changes, people walk into the robot's path. A cup can be full, fragile, wet, or unexpectedly heavy. Success depends not only on recognizing the cup, but on applying the right forces at the right moment. A wrong answer from a chatbot might be embarrassing. A wrong movement from a heavy machine might break someone's hand. Robot videos rarely show the denominator. We see the successful attempt, but not always how many failed attempts came before it. Was the machine autonomous throughout? Was the room rearranged beforehand? Did a human remotely rescue it? How quickly could staff restore it to service? These are not hostile questions. They are the ordinary questions any serious buyer must ask. And this brings us to one of the biggest misunderstandings in robotics. The belief that walking is the hard part and useful work comes naturally afterward. Walking is difficult, but manipulation is often even harder. Human hands are extraordinarily sophisticated. We can pull a thin sheet from a stack, thread a needle, fold fabric, and rotate a key without consciously calculating the physics. Our sense of touch tells us when an object starts to slip. Robots need hardware and control systems that approximate those abilities, often under tight weight, power, and cost constraints. A robot capable of lifting a heavy box may struggle with an ordinary crumpled shirt. Factories reduce that difficulty by making the environment predictable. Objects arrive in specific containers. Floors are mapped. Tasks can be separated into repeatable cycles. Specialized grippers can replace generalpurpose hands. Engineers can surround the robot with additional sensors or safety equipment. That is not cheating. It is how industrial automation becomes reliable. But it explains why a successful warehouse robot does not automatically become a capable household assistant. What about the economics? Imagine a factory evaluating a humanoid with a hypothetical purchase price of $50,000. That figure is an example for calculation, not a verified market price for all the robots we have discussed. Spread over 5 years, the hardware alone averages $10,000 per year before financing or depreciation details. Add maintenance, replacement parts, electricity software connectivity safety equipment, integration, and human supervision. Then account for downtime, and the fact that a robot may work more slowly than a trained employee. A robot that replaces one repetitive task in a high wage operation across long shifts might justify that investment. The same machine used for only an hour or two each day may not. A useful financial comparison is cost per correctly completed task, including the cost of failures and supervision. Some customers may prefer a robot as a service contract instead of buying the machine outright. That can reduce upfront spending, but it does not eliminate the underlying economics. A striking factor is utilization. If a robot can do one job for 10 hours daily, its price is spread across thousands of productive hours. If it spends most of its life idle in a home, the effective cost of each useful action becomes much higher. This is why warehouses and factories may be the first sustainable markets. Even though a domestic robot is the image that captures everyone's imagination, and the home is a much harsher test than a factory, consider a single request. Please clean the kitchen. A person understands that this might involve putting food away, distinguishing dirty dishes from clean ones, navigating around a pet, wiping a counter without knocking over a glass, and deciding not to throw away something important. The home contains fragile objects, unpredictable children, pets, stairs, cords, spills, and almost unlimited variation. Every household has different furniture, different routines, and different definitions of a tidy room. Safety standards are another obstacle. A factory can establish marked work zones, trained personnel, and emergency procedures. A robot in a home must operate near sleeping people, vulnerable relatives, and unsupervised children. It may need privacy preserving cameras and microphones. It must protect the data it collects. Its owner must be able to stop it reliably, maintain it, and understand who is responsible when it damages property or injures someone. There are also mundane questions. Will it fit through the doorway, charge itself, survive a spilled drink, and receive years of software updates? So, when might we realistically have humanoid robots at home? There is no verified date when a safe, general purpose, and affordable household humanoid will be widely available. Some companies have suggested aggressive timelines, but forecasts should not be mistaken for delivery schedules. Limited home trials and expensive early adopter systems are more plausible in the nearer term than a robot that can reliably handle most domestic chores without assistance. Even a relatively inexpensive consumer humanoid would be hard to justify financially if it could perform only a few carefully rehearsed tasks. Broad adoption requires several breakthroughs to occur together. Hardware must become much cheaper to manufacture and repair. Hands must improve. Batteries and power management must support useful working time. Software must generalize across unfamiliar homes rather than memorizing demonstrations. Failure rates must fall dramatically. Safety certification, insurance, cyber security, and long-term servicing need to become routine. And most importantly, the robot must save enough human time to justify its total cost. It is entirely possible that the first genuinely useful home robots will not be full-size humanoids. They may be smaller, slower, wheeled machines designed for limited tasks. A lowcost robot that safely performs three chores every day could be more valuable than an impressive biped that can do 30 tricks on stage. There is also a social question that numbers alone cannot settle. If a robot performs warehouse tasks, how does the employer retrain and protect existing workers? If it enters a home to help an older person, how do we make sure companionship and care do not become excuses for removing human contact? And if the machine is constantly observing its environment to learn who owns the resulting images and recordings, none of these questions means humanoid robots are a fantasy. The strongest evidence now points in the opposite direction. Carefully designed robots are carrying out real material handling jobs and customers are reporting measurable outputs. The breakthrough is not that a robot can dance or that it can answer a question or that it can briefly imitate a human hand. It is that a physical machine can become sufficiently dependable to perform a defined job in the imperfect conditions of a working business. But the distance from that achievement to an all-purpose domestic assistant remains enormous. Perhaps the best way to think about the future is not to ask which company has the most humanl looking robot. Ask three much less glamorous questions. How often does it succeed? How much does every successful job actually cost? And how much help does it still need from people? The winners in embodied AI may not be the robots with the most spectacular videos. They may be the ones whose work becomes so ordinary that nobody bothers to film it anymore.

Assistir ao vídeo original no YouTube

Assuntos: Robótica e humanoides, Robótica e humanoides. Publicado no YouTube em 2026-10-25.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.

About Honda's path from the E0 to ASIMO

Honda's official robotics history starts in 1986 with the E0, a robot built to study the principle of walking on two legs. The first model with a torso and arms, the P1, was 1,915 mm tall, weighed 175 kg and used an external power source and computer; it could turn a switch on and off, grab a doorknob and carry objects. The robot Honda presented in December 1996 as the world's first bipedal humanoid walking robot was 1,820 mm tall and weighed 210 kg, with computers, motor drives, batteries and wireless equipment placed inside the torso so it could work without cables. In September 1997 Honda completed a fully independent version that was 1,600 mm tall and weighed 130 kg.

ASIMO was announced in 2000, smaller and lighter and using the i-WALK walking technology, designed for ordinary living spaces, including stairs and slopes. According to Honda, a version for the rental business followed in 2001; in 2002 ASIMO gained intelligence technology that let it interpret people's postures and gestures, greet and follow people, recognize faces and address people by name, and provide information over a network while doing reception work. In 2004 Honda announced technologies for faster, continuous and autonomous movement, including human-like running, moving in sync with people, pushing a cart and carrying trays.

Fact sheet

• Company: Honda Motor Co., Ltd.
• First experimental robot: E0 (1986), for studying bipedal walking
• First model with torso and arms: P1: 1,915 mm and 175 kg
• Robot presented in December 1996: 1,820 mm and 210 kg, cable-free
• Robot completed in September 1997: 1,600 mm and 130 kg
• ASIMO announced: 2000, with i-WALK technology
• Later updates: 2001 rental version, 2002 intelligence technology, 2004 running and cart operation

Read also

• The people filming their chores to train humanoid robots
• Why $55,000 Humanoid Robots Still Struggle With Housework
• Can Humanoid Robots Take Over Supermarket Shelves? | Tech It Out

---

Sources and references: Honda – History of robotics development · video by the Daily AI Snapshot channel on YouTube.