Wall Street's conviction around humanoid robots entered a new phase this week. In an interview with CNBC published Wednesday, Wedbush Securities managing director Dan Ives called the category "the golden goose for physical AI" and said Wedbush sees it as "one of the biggest market opportunities in the AI Revolution." Separately, SoftBank CEO Masayoshi Son told CNBC on June 1 — one day after SoftBank announced a €75 billion investment in AI infrastructure in France — that humanoid and industrial robotics, "with physical AI as a core," is where he expects the next trillion-dollar company to emerge.
Both statements land against a data point that has not softened. Goldman Sachs revised its 2035 humanoid robotics market forecast upward from $6 billion to $38 billion — a sixfold increase — citing AI advances and accelerating cost reduction as the reasons. The bank's research team points specifically to end-to-end AI training, which allows models to train themselves without manual engineering of every behavior, as the single development that most exceeded their earlier expectations. Goldman's base case now calls for more than 250,000 humanoid robot shipments in 2030, almost entirely for industrial use.
Physical AI Market Shows Commercial Deployment Milestones Are Real
For institutional investors, the shift from forecast to hard evidence has been the defining development of 2026. Agility Robotics' Digit has moved more than 100,000 totes at a GXO Logistics facility since its first commercial deployment in June 2024. In February 2026, Agility signed a Robots-as-a-Service agreement with Toyota Motor Manufacturing Canada following a successful year-long pilot; seven Digit units are now handling material logistics at the Woodstock, Ontario plant where Toyota builds the RAV4.
Boston Dynamics' electric Atlas has had its entire 2026 production allocation committed, with first deployments going to Hyundai, where the robot has been filmed sorting car parts in a live production setting. Figure AI's Figure 02 completed an eleven-month deployment at a BMW plant in South Carolina, during which two robots assisted in assembling more than 30,000 cars and handled more than 90,000 sheet metal parts. The supply-side constraint is itself a market signal: for at least one leading platform, demand has already outrun what manufacturers can currently build.
How Sim-to-Real Training Cuts Costs From $150K to $40K
The technical architecture that makes the current cost compression possible is not a manufacturing trick — it is a software paradigm shift. Companies including Agility Robotics, Figure AI, and NVIDIA have converged on a training method called sim-to-real transfer: a robot's entire behavioral repertoire is learned in a physics simulator, then transferred to physical hardware with no additional tuning required.
Agility Robotics' whole-body control system for Digit is an LSTM-based neural network with fewer than one million parameters, trained in NVIDIA's Isaac Sim for what amounts to decades of simulated experience over three to four days of compute time. The resulting policy transfers zero-shot to physical hardware — meaning the robot can perform a task in the real world immediately after training in simulation, without further adjustment. Figure AI uses a reinforcement learning approach in which thousands of virtual robots with randomized physical parameters are trained simultaneously, with the policy then applied directly to factory hardware.
NVIDIA's GR00T N1 model operates as a Vision-Language-Action architecture — a dual-system design in which a vision-language module interprets the environment and a diffusion transformer module generates continuous motor commands in real time, both trained end-to-end on a mixture of real robot trajectories, human video, and synthetically generated data. This approach enables a single model to generalize across different robot bodies and task types, which is why NVIDIA can supply it as infrastructure rather than requiring each manufacturer to build behavioral models from scratch.
The economics that follow from this architecture are direct. The largest cost components in a commercial humanoid are actuators — approximately 35 to 40 percent of the bill of materials — batteries at 15 to 20 percent, and onboard compute at 10 to 15 percent. Chinese manufacturers, particularly Unitree, have driven actuator costs down by approximately 50 percent through domestic production at scale, and battery costs have fallen to roughly $100 per kilowatt-hour in 2026. The combination of commoditized hardware and simulation-trained software is what moved the commercial-grade price band from $150,000 and above in 2023 to approximately $40,000–$60,000 in 2026. Agility's Robots-as-a-Service rate at Toyota — approximately $30 per hour per robot — bundles hardware, software, over-the-air model updates, and maintenance into a single operational line item, allowing manufacturers to adopt the technology without upfront capital expenditure.
China Leads on Volume While Unitree Security Flags Mount
The competitive picture is not straightforward for Western investors. Wedbush's Ives told CNBC directly that China is currently the clear leader in humanoid robotics, with the U.S. playing catch-up. Unitree Robotics, headquartered in Hangzhou, shipped approximately 5,500 units in 2025 and is targeting 10,000 to 20,000 in 2026. Chinese companies accounted for a majority of 2025 global humanoid installations, reflecting an early-mover advantage that Western competitors — still largely in pilot phases through most of that year — had not yet closed.
That volume advantage now carries structural complications. On June 1, NVIDIA announced that its first publicly available humanoid robotics research system — combining Unitree's H2 chassis with NVIDIA's Jetson Thor hardware and Isaac GR00T software — would be made available to academic research labs including Stanford University and the University of California, San Diego. The announcement prompted immediate scrutiny because Unitree is a Chinese company subject to Article 7 of China's National Intelligence Law, enacted in June 2017, which requires every Chinese company and citizen to support, assist, and cooperate with state intelligence work. That obligation applies regardless of where the company's servers are located, where its subsidiaries are incorporated, or what its privacy policy states. It is not a risk to weigh against performance — it is the operative law of Unitree's home jurisdiction.
No Independent Audit Has Cleared Unitree Humanoid Models
Congressional concern over Unitree predates the NVIDIA partnership. In May 2025, a bipartisan group of 24 U.S. Representatives urged the Defense Secretary, Commerce Secretary, and FCC Chair to investigate Unitree's ties to People's Liberation Army-affiliated institutions and consider adding the company to federal security designation lists. The House Homeland Security Subcommittee held a dedicated hearing on Unitree's national security risks on March 17, 2026. Senators Rick Scott and Tom Cotton introduced the Blocking CCP Spy Tech Act of 2026 in May, which names Unitree specifically and would require a national security investigation that could result in the company being added to the FCC's Covered List — a designation that would restrict its use in telecommunications infrastructure and could affect researchers receiving federal funding.
Security researchers confirmed a documented backdoor, designated CVE-2025-2894, in Unitree's Go1 robot dogs: a hidden CloudSail tunnel service that automatically connected to Unitree's servers, giving anyone with the management key access to the robot's camera, audio, and sensor data. Unitree deactivated the service and stated that newer models, including its humanoid robots, have a more secure architecture. No independent security audit of the humanoid models has been published confirming that assessment, and security researcher Víctor Mayoral-Vilches of Alias Robotics found separate undisclosed streaming of telemetry — including audio, visual, and spatial data — to servers in China in the G1 humanoid specifically. Researchers also found that Unitree's implementation relies on hardcoded encryption keys that had already leaked online. Unitree has denied wrongdoing.
NVIDIA addressed these concerns directly and told Reuters that all software updates for the robot's subsystems will flow through its Blackwell chip, where code is authenticated before execution. NVIDIA also confirmed plans to build similar reference platforms with humanoid manufacturers in the U.S., Europe, and South Korea — partners it declined to name publicly.
What Are the Real Engineering Limits of Today's Humanoid Robots?
Goldman Sachs, despite revising its forecast upward, was explicit in identifying where the technology falls short. The bank flagged robot manipulation — grasping objects of varying shape, weight, and compliance — and natural interaction with voice commands as the two remaining bottlenecks that mass-produced general-purpose humanoids have not solved. Current platforms excel in structured environments where tasks are repetitive and surfaces are predictable. They struggle in unstructured settings.
Battery life imposes a practical ceiling: current-generation commercial humanoids achieve two to eight hours of continuous operation before requiring a recharge, a constraint that limits applicability to environments where downtime is costly. Locomotive reliability on novel surfaces remains unproven outside of purpose-designed factory floors. Expansion joints wider than 10 millimeters can trip bipedal robots; oil residue on concrete creates fall risk that ANSI/A3 R15.06-2025, the current U.S. safety standard for industrial robots, was not designed to address. ISO 25785-1, the walking-robot standard that would cover fall zone calculations for bipedal systems, remains a working draft and has not been published.
There is also a documented safety gap specific to force output. In November 2025, a federal whistleblower lawsuit filed in the Northern District of California alleged that Figure AI's F.02 humanoid was producing impact forces that internal testing found to be more than twice those needed to fracture an adult human skull — and that the company fired the engineer who raised those concerns. Figure AI denied the allegations. The case is pending. No OSHA standard specifically addresses humanoid robot force output; the agency applies its General Duty Clause and refers employers to ANSI consensus standards, none of which was written for bipedal walking machines.
For investors, analysts note that roughly 25 percent of warehouse labor currently involves tasks too variable for conventional fixed automation — the category humanoid robots are best positioned to absorb. The World Economic Forum's 2025 Future of Jobs Report projected a net gain of 78 million jobs globally by 2030, with 170 million new roles created and 92 million displaced. McKinsey Global Institute's range is wider and less optimistic, with automation potentially displacing between 400 million and 800 million jobs globally by 2030. Neither estimate resolves the transition-period question: who bears the cost of displacement in the years before new roles materialize.
The consensus among analysts is that 2026 marks the market's genuine beginning, not its maturation. Goldman Sachs's estimate of $38 billion by 2035 reflects a field where physical hardware has become commercially viable, AI training pipelines have reached industrial scale, and the leading volume supplier happens to sit inside a jurisdiction whose law requires it to cooperate with state intelligence. For any enterprise or institution evaluating humanoid robots today, the investment case is real, the engineering limits are specific, and the Chinese market leader arrives with a legal obligation that no privacy policy can override.
Frequently Asked Questions
How do humanoid robots learn new tasks in 2026?
Most commercial humanoid robots in 2026 learn tasks through sim-to-real transfer: the robot's control policy is trained in a physics simulator — such as NVIDIA's Isaac Sim — through reinforcement learning over thousands of virtual hours, then deployed on physical hardware without additional tuning. NVIDIA's GR00T N1 foundation model adds a vision-language module that allows robots to interpret natural language instructions and generalize across different robot bodies and tasks.
What is the real cost of deploying a humanoid robot in 2026?
Commercial-grade humanoids currently cost between $40,000 and $60,000 per unit, down from over $150,000 in 2023, driven by cheaper actuators and battery packs. Companies including Agility Robotics offer a Robots-as-a-Service model at approximately $30 per hour per robot, which bundles hardware, software, and maintenance into a single operating cost roughly comparable to the fully loaded cost of a human worker in a high-wage manufacturing region.
Is it safe to use Unitree robots in U.S. facilities?
Unitree's Go1 robot dogs were confirmed to contain a hidden backdoor (CVE-2025-2894) that transmitted data to servers in China; Unitree deactivated the service and says newer models are more secure, but no independent audit of its humanoid models has been published. Under Article 7 of China's National Intelligence Law, Unitree is required to cooperate with Chinese state intelligence regardless of where its data is stored or where it operates. A pending Senate bill — the Blocking CCP Spy Tech Act of 2026 — would require a national security investigation and could restrict federally funded researchers from using Unitree hardware.
What is the Goldman Sachs humanoid robot forecast for 2035?
Goldman Sachs Research revised its 2035 humanoid robot market forecast upward to $38 billion — six times its original $6 billion estimate — citing AI advances that allowed models to train themselves and falling manufacturing costs. The bank's base case calls for more than 250,000 humanoid robot shipments in 2030, with demand concentrated in structured industrial environments such as automotive manufacturing and logistics.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
