In 2025, discussions about robotics increasingly revolve around specific names — much like the automotive or semiconductor industries. Atlas from Boston Dynamics and Optimus from Tesla have become symbols of two poles in the evolution of humanoid robots: the former pushing the limits of dynamics and bodily control, the latter designed from the outset as a scalable industrial tool for the economy. Between these extremes, an entire layer of companies — from Figure and Agility Robotics to UBTECH and Sanctuary AI — is rapidly emerging, seeking to merge both paradigms: physical expressiveness and direct integration into production and service chains.
The following is an overview of what the leading humanoids look like today, who has truly brought robots into paid employment, and what the embodied AI market might look like over the next 10–15 years.
Atlas vs. Optimus: Two Poles of the Same Future
The new fully electric Atlas, unveiled by Boston Dynamics in 2024, is officially described as “the most dynamic humanoid in the world.” It is no longer a laboratory prototype but a platform designed for real industrial tasks. The company highlights its transition from hydraulics to a fully electric drivetrain, a more compact mechanical layout, and a focus on “whole-body mobility.” The robot can rotate its limbs far beyond the human range, perform powerful twisting motions and extreme torso rotations, and rise from complex poses — demonstrating movements that traditional manipulators cannot replicate.
This represents the logic of expanding physical boundaries. Atlas functions as an experimental platform for Hyundai and its industrial partners, testing how far balance, stability, jumping dynamics, and high-speed interaction with objects can be pushed before entering mass production. It is the “Formula One car” of the humanoid world, where cutting-edge solutions are tested before being simplified for mainstream machines.
Optimus, on the other hand, approaches from the opposite end. Tesla describes it as a full-size humanoid roughly 173 cm tall and weighing around 70–75 kg, with over 40 degrees of freedom, including 11 degrees of freedom per hand and integrated tactile sensors. The geometry is intentionally modeled on average human proportions: the robot is meant to work in existing environments without modifications, reaching the same shelves, handling the same tools, and lifting the same boxes as human workers.
The key turning point in 2025 came when Tesla abandoned motion-capture suit training in favor of video-based learning, applying the same “vision-only” philosophy used in its Full Self-Driving (FSD) system. The company is now training humanoids directly on factory video footage of human workers. Cameras record actual operations, and neural networks learn to reproduce and adapt those behaviors in specific production settings. In this logic, the factory transforms from a programmed space into a learning environment: humans perform tasks once, the dataset grows, and the robot gradually assumes increasingly complex operations.
In essence, Atlas represents the maximization of physical expression and control, sometimes beyond human physiology, while Optimus embodies maximum integration into existing economic and AI ecosystems. Atlas shows how far the body can be mastered; Optimus shows how deeply that body can be embedded into industrial and logistical systems. Together, they form the two ends of a much broader landscape.
The Third Force: Figure, Agility, Apptronik, and Sanctuary
Over the past two years, a strong new layer of independent players has grown around Tesla and Boston Dynamics, quietly advancing humanoid robots into the realm of industrial routine.
Figure AI is building its Figure 02/03 line as a family of multi-purpose humanoids: first for factory work, then for domestic environments. The company publicly announced a collaboration with BMW, revealing that Figure 02 had already operated on the assembly line at the Spartanburg plant, taking part in assembling approximately 30,000 vehicles — effectively becoming the first general-purpose humanoid to perform regular paid work in the automotive industry. The next generation, Figure 03, is being reimagined as a home assistant, not just a factory worker.
The key to Figure’s approach lies in its proprietary AI system Helix, a vision–language–action model that governs perception, motion, and reasoning onboard in real time. According to the company, this allows the robot to plan and execute tasks without rigid scripting. It is an attempt to make the status of “general-purpose robot” not a marketing slogan, but an actual architectural reality, where body and mind are designed as a single, integrated system.
Agility Robotics, with its bipedal Digit, is betting on a more utilitarian niche — logistics and warehousing. Digit, standing about 1.75 meters tall and capable of carrying loads of up to 16 kilograms, has already been declared the “first commercially deployed humanoid robot.” Companies like Amazon and several logistics operators are testing Digit for package sorting and last-mile delivery tasks. Agility is doing what Atlas and Optimus are still approaching — selling fleets of humanoids as ready-made B2B solutions, complete with docking stations, auto-charging, and a library of pre-trained skills for standard warehouse operations.
Apptronik, creator of Apollo, emerged from the Human Centered Robotics Laboratory at the University of Texas and is now testing its robots on Mercedes-Benz production lines and with contract manufacturer Jabil. Apollo, roughly human-sized and weighing about 160 pounds, can lift up to 25 kilograms and work for up to 22 hours a day with periodic charging. Its design philosophy is rooted in compatibility with existing workspaces — robots should fit into current factory layouts, not force costly redesigns. In 2025, the company raised hundreds of millions of dollars and reached a valuation of around $5 billion, making it one of the strongest contenders for the title of “Tesla analog in the humanoid niche.”
Sanctuary AI, based in Vancouver, is following an even more cognitive path. Its Phoenix humanoid is positioned as a general labor robot, powered by the Carbon AI system, which aims to achieve a “human-like intelligence level” for hundreds of professional tasks across retail, logistics, and services. The company has already demonstrated the robot performing hundreds of client-specific operations and has attracted strategic investors such as Accenture, who explicitly view Phoenix as the future embodiment of both white- and blue-collar work in a single form.
Finally, 1X Technologies, partially backed by OpenAI, extends the humanoid concept beyond factories. Its NEO model is a domestic assistant, initially operating under a hybrid teleoperation model: in complex scenarios, a remote human operator takes control, executes an action, and thereby teaches the model new skills. It is a pragmatic compromise — before vision–language–action (VLA) systems become reliable enough to manage an apartment entirely autonomously, part of the intelligence will remain human-in-the-loop, ensuring safety and accelerating training.
The Chinese Factor: Walker S2 and the Explosive Growth of the Market
In parallel, China is building its own large-scale testing ground for embodied AI. With its industrial Walker S2, UBTECH demonstrated in 2025 what could become the first true “24/7 worker”: a full-size humanoid robot with 52 degrees of freedom, capable of autonomously swapping its own batteries in three minutes. The robot approaches the station, removes a depleted module, inserts a charged one, and resumes operation — all without human involvement.
This solution establishes the architecture for round-the-clock robotic shifts — continuous operation without downtime and without complex fast-charging infrastructure. UBTECH has already begun mass production of hundreds of these robots and plans to scale up deliveries, while simultaneously signing contracts for hundreds of Walker S2 units to patrol border crossings in southern China. Thus, the Chinese humanoid is entering not only factories, but also security infrastructure — from border control to urban surveillance.
Against this backdrop, officials in Beijing are publicly warning about the risk of a market bubble: more than 150 companies in China claim to be developing humanoid robots, and the pace of investment is now outstripping the actual maturity of the market. Yet this very overheating is generating a massive portfolio of technologies, from which several global champions will inevitably emerge — especially if they succeed in linking local hardware expertise with global AI stacks.
From Hardware to Brain: Nvidia Isaac GR00T and the Era of Foundation Models for Robots
If Atlas, Optimus, Figure, and their counterparts have solved the “body” problem, the next frontier is the shared “brain” for the entire class of humanoid machines. In 2025, NVIDIA launched Isaac GR00T N1, followed by an upgraded GR00T N1.5 — open foundation models for humanoid robots, described as vision–language–action systems trained on a mix of egocentric human videos, robotic trajectories (both real and simulated), and synthetic datasets.
According to NVIDIA, GR00T N1.5 is a cross-embodiment model that takes language and images as inputs and generates motor action sequences. Its architecture explicitly separates a “fast” reflexive layer from a “slow” planning module, mirroring the two-system model in cognitive psychology. The model is already being demonstrated on real humanoid platforms, including Fourier’s GR-1, and integrated into the future “reference factories” of Foxconn in the United States — facilities that will assemble NVIDIA servers using humanoids running under the GR00T operating stack.
At the GTC conference, Jensen Huang framed this as the beginning of the “era of generalist robotics” — a world in which a single foundation “brain” can operate across Atlas-like, Optimus-like, Walker S2-like, and domestic Neo-like machines, with differences reduced to adaptation layers and mechanical calibration.
Futuristic Outlook: How the Market Might Look by the Mid-2030s
When all these vectors are combined — the physical platforms, new learning models, and foundation systems like GR00T and Helix — a fairly clear picture emerges for the 2030–2035 horizon.
First, humanoids are set to become a standard option in the automotive, logistics, and light manufacturing industries. Already today, Figure 02 and Digit are physically performing shifts at BMW factories and in Amazon’s logistics centers; Apollo is entering pilot programs with Mercedes-Benz and Jabil; and Optimus is expected to be deployed at Tesla factories toward the end of the decade. Within the next five to seven years, this could evolve into a landscape in which every major industrial plant employs a “layer” of humanoids, handling 10–30% of routine operations currently performed by humans or traditional robots.
Second, a clear stratification will take shape. Atlas-like platforms will remain the Formula One prototypes for high-dynamics tasks — disaster response, heavy construction, emergency work. Optimus-type robots will occupy the mass-labor tier in standardized environments: assembly, logistics, and simple service operations. Walker-type machines, with autonomous battery swapping, will underpin 24/7 security and infrastructure networks — from border control to large transportation hubs. Meanwhile, Figure- and Phoenix-class humanoids will compete for the role of multi-purpose employees, capable of being retrained from warehouse duties to retail or light assembly.
Third, humanoids will enter homes through a hybrid model. The first generations of Figure 03 and 1X NEO domestic robots will likely operate in a mixed mode — partly teleoperated, partly autonomous. Human operators will remain the invisible layer between the user and the foundation model, preventing over-expectations from AI while simultaneously accumulating vast, real-world datasets for the fully autonomous generations that will follow.
Fourth, the regulatory and ethical front will become as crucial as the technological one. China is already experimenting with humanoid robots in border and policing infrastructure, while UBTECH is signing multi-million-dollar contracts for Walker S2 units. Within the next decade, we are likely to see international agreements restricting the use of humanoids in combat and repressive scenarios, similar to ongoing discussions about autonomous weapons systems.
Finally, the economic question looms largest. Analysts and industry leaders are already warning of a potential bubble: valuations and funding volumes for startups such as Figure, Apptronik, and dozens of Chinese firms are growing faster than the actual volume of paid, real-world tasks available for humanoids. Some projects will inevitably fail during the first wave of disillusionment. Yet, much like the dot-com era of the early 2000s, the infrastructure created during this period of overheating — the models, datasets, manufacturing chains, and safety standards — will remain and form the foundation for the next, more mature wave.
Robots of War and Disaster: The Dark Side of Embodied AI
While humanoids are learning to move boxes and work on assembly lines, another flank of robotics is evolving in far less comfortable environments — on battlefields and in disaster zones. Here, the focus is not on “human likeness” per se, but on survivability, autonomy, and the ability to operate where humans cannot or should not go.
In the defense sector, a whole family of unmanned ground vehicles (UGVs) has already moved beyond trade-show exhibits and into real testing grounds. These include tracked infantry-support robots, heavy combat UGVs, and autonomous logistics platforms. They can perform reconnaissance, advance under fire, evacuate the wounded, deliver ammunition, and — in certain configurations — carry remotely controlled weapons modules.
The technological stack behind them is essentially the same as in civilian autonomous systems: lidar, cameras, radar, SLAM navigation, network protocols that allow swarm coordination, and an ever deeper integration with AI-driven object recognition and path-planning systems. The crucial difference is that an algorithmic error in combat doesn’t mean lost productivity — it means human casualties. This alone has already triggered an entire field of ethical and legal debate surrounding such systems.
At the other end of the spectrum are robots designed for disaster response and industrial accidents. Here, the demands are almost the opposite of military applications: instead of speed and firepower, the priorities are resistance to radiation and chemical contamination, mobility in debris, and the ability to reach areas inaccessible to humans or traditional manipulators. Snake-like robots crawl through collapsed tunnels and buildings, quadrupeds and tracked platforms equipped with thermal cameras search for survivors under rubble, while lightweight drones map the destruction in three dimensions, helping rescue teams plan operations in real time.
Humanoid platforms — especially the “dynamic” descendants of Atlas — are increasingly becoming the universal prototypes for the next generation of search-and-rescue robots. Where today an entire fleet of specialized machines is required, in the future it may be possible to deploy a single multi-purpose humanoid system capable of clearing debris, connecting equipment, and acting as a human-tool interface.
Both military and rescue robots rest on the same embodied AI foundation, and this creates the classic dual-use dilemma: every improvement in sensors, actuators, batteries, or foundation models simultaneously makes humanoids more useful in logistics and more effective in combat. As a result, a race of norms is now unfolding in parallel with the race for hardware — from UN debates on autonomous weapons to industry codes of conduct defining unacceptable uses of AI.
Instead of a Conclusion: Who Will Win the Race
Looking at the landscape of late 2025, it’s clear that the question is no longer whether humanoid robots will work alongside humans — but who will define the standards for this coexistence. Boston Dynamics, with Atlas, will continue to set the physical limits of what a bipedal, two-armed machine can achieve. Tesla, with Optimus, aspires to be the platform that transforms the factory itself into a training ground for embodied AI.
Figure, Agility, Apptronik, Sanctuary, and UBTECH are already embedding humanoid robots into real industrial and service workflows, while NVIDIA and other foundation-level players are striving to make it possible for any manufacturer to “plug in a brain” to its hardware without years of in-house AI development.
By the 2030s, it is highly likely that we will no longer be discussing one-off demonstrations but facing a classic market question: Which manufacturer and which “brain” do you choose for your factory, logistics center, clinic, or residential complex? The current race for attention, funding, and technological dominance will be remembered as only the first act in a much larger story — a story about how humanity negotiates with millions of new “working bodies” endowed with artificial minds, becoming an integral part of our daily reality.
At the same time, a new military generation of embodied systems is emerging, built on the same foundations — and the decisive question will be whether these systems are capable or incapable of guaranteeing human security on a global scale. That will be the true test of the century.
