The Humanoid Robot Race: How AI Is Powering the Next Generation of Physical Automation

The Year Physical AI Gets Real

For decades, humanoid robots existed primarily in science fiction and carefully choreographed demo videos. That era is over. In 2025-2026, the humanoid robotics industry has undergone a phase transition, with billions of dollars in investment, commercial deployments in real factories, and a convergence of AI capabilities that makes general-purpose humanoid robots genuinely feasible for the first time.

Figure AI, the Sunnyvale-based startup, raised $675 million in a Series B round at a $2.6 billion valuation in February 2024, with investors including Microsoft, Nvidia, Jeff Bezos, Amazon, and OpenAI. By September 2025, the company had closed a Series C round exceeding $1 billion at a staggering $39 billion post-money valuation — a 15x increase in just 18 months, signaling extraordinary investor confidence. The company’s Figure 02 robot completed an 11-month deployment at BMW’s Spartanburg, South Carolina manufacturing plant, running 10-hour shifts five days a week, loading over 90,000 parts, and contributing to the production of more than 30,000 BMW X3 vehicles with accuracy above 99%. In January 2026, Figure AI unveiled its third-generation Figure 03, featuring tactile fingertip sensors that detect forces as small as three grams, cameras with double the frame rate and a quarter of the latency of its predecessor, and wireless inductive charging through the robot’s feet.

Tesla’s Optimus program tells a more complicated story. Despite ambitious targets of 5,000 units for 2025, actual production fell far short, with only hundreds of units manufactured before a design pause. In January 2026, Elon Musk admitted that no Optimus robots were yet doing useful work in Tesla’s factories. The company is now preparing a V3 prototype for Q1 2026, with production-intent units targeted before year-end and a dedicated factory at Giga Texas planned to eventually produce up to 10 million units annually. Boston Dynamics’ fully electric Atlas, unveiled in April 2024, has moved faster toward commercialization: the company began commercial production in early 2026, with all deployments already committed to Hyundai Motor Group manufacturing facilities and Google DeepMind.

The investment landscape tells a compelling story. Goldman Sachs has projected the humanoid robot market could reach $38 billion by 2035 — a sixfold increase from its earlier $6 billion estimate — with shipments forecast at 1.4 million units. Morgan Stanley’s more aggressive outlook projects $357 billion by 2040 and a $5 trillion total addressable market by 2050, implying roughly one billion humanoids in service. In 2025, robotics startups pulled in over $6 billion in venture capital globally, with humanoid companies capturing a major share: Figure AI alone raised over $1 billion, Apptronik secured $403 million, and NEURA Robotics closed a EUR120 million Series B. This is not speculative enthusiasm — it is a calculated bet that the convergence of large language models, computer vision, and advanced actuator technology has finally made humanoid robots economically viable for industrial work.

The Convergence: LLMs Meet Physical Bodies

The technical breakthrough driving the humanoid revolution is not hardware — it is AI. Specifically, the integration of large language models and foundation models with robotic control systems, a paradigm researchers call “embodied AI.” The insight is straightforward: the same transformer architectures that power ChatGPT can be adapted to process sensory inputs (camera feeds, tactile sensors, force measurements) and output motor commands.

Google DeepMind’s RT-2 (Robotics Transformer 2), introduced in July 2023, demonstrated that vision-language models could directly control robotic arms, translating natural language instructions into physical actions without task-specific programming. RT-2 improved performance on novel scenarios from 32% to 62% compared to its predecessor, and could perform multi-stage semantic reasoning — such as identifying which object could serve as an improvised hammer. This approach has since been scaled by multiple companies. Figure AI initially partnered with OpenAI in 2024 to enable its robots to understand spoken instructions and reason about their environment. However, Figure ended that partnership in 2025, citing that large language models were becoming commoditized, and announced a major breakthrough in fully end-to-end robot AI built entirely in-house. Figure 03 now runs on the company’s proprietary Helix AI system, a vision-language-action model that processes camera feeds, understands spoken commands, and generates complex task plans without external AI dependencies.

Nvidia’s Project GR00T (Generalist Robot 00 Technology), announced at GTC 2024, provides a foundation model specifically designed for humanoid robots, enabling them to learn from observing human demonstrations. The platform processes multimodal inputs — text, video, sensor data — and generates coordinated whole-body movements. At GTC 2025, Nvidia went further, releasing Isaac GR00T N1 as the world’s first open, fully customizable foundation model for humanoid reasoning. GR00T N1 uses a dual-system architecture: a vision-language module interprets the environment, and a diffusion transformer module generates fluid motor actions in real time. The model was trained on a heterogeneous mix of real-robot trajectories, human videos, and synthetic datasets, and has been released as open source on Hugging Face, with early access partners including Agility Robotics, Boston Dynamics, Mentee Robotics, and NEURA Robotics. Combined with Nvidia’s Isaac Sim for photorealistic simulation, the development cycle for new robotic behaviors has compressed from months to days.

Manufacturing and Warehouse Deployments: Where the Money Is

The initial commercial market for humanoid robots is concentrated in two sectors: automotive manufacturing and warehouse logistics. These environments are structured enough to be tractable for current robotic capabilities, yet complex enough that traditional fixed automation (robotic arms, conveyor systems) cannot handle every task.

In automotive manufacturing, humanoid robots address the “last mile” of automation. Modern car factories are already 80-90% automated for welding, painting, and major assembly operations. But final assembly tasks — threading wire harnesses, installing interior components, performing quality inspections in confined spaces — remain manual because they require human-like dexterity and spatial reasoning. Figure AI’s completed deployment at BMW demonstrated that its robots could maintain cycle times of 84 seconds per task (including 37 seconds for the actual load) with accuracy above 99%, operating on 10-hour shifts. The deployment also surfaced real engineering challenges: forearm components proved to be the top hardware failure point, leading to a complete wrist redesign for Figure 03. With the new Figure 03 fleet now deploying at BMW in early 2026, early reports suggest a significant jump in operational efficiency. Meanwhile, Boston Dynamics has committed its entire 2026 Atlas production run to Hyundai’s Robotics Metaplant Application Center, and Apptronik’s Apollo has secured commercial agreements with both Mercedes-Benz and GXO Logistics for pilot deployments in manufacturing and logistics.

The warehouse story is more nuanced than initial hype suggested. Amazon now operates over one million robotic units across its fulfillment network — surpassing the 750,000 milestone from 2023 — but these remain specialized machines (Proteus mobile robots, Sparrow picking arms) handling discrete tasks. The company’s experiment with general-purpose humanoids has hit a significant wall: Amazon cancelled its Digit deployment program with Agility Robotics after the bipedal robots failed to meet Amazon’s exacting operational standards. This is a sobering data point for the industry. That said, Agility Robotics has found traction elsewhere — its Digit robots have moved over 100,000 totes at GXO’s Flowery Branch facility, demonstrating that warehouse humanoid deployment can work in the right context, if not yet at Amazon’s scale and speed requirements.

The China Factor: A New Front in the Robot Race

One of the most significant developments in 2025-2026 that reshapes the competitive landscape is China’s rapid emergence as a humanoid robotics powerhouse. Chinese firms accounted for roughly 80% of the approximately 13,000 humanoid robots shipped globally in 2025. Unitree Robotics shipped over 5,500 humanoid units in 2025 alone, surpassing all US-based peers, and has set a target of 20,000 units for 2026.

The pricing implications are enormous. Unitree launched its G1 humanoid at $16,000 in 2024, then released the even more affordable R1 at $5,900 in mid-2025 — prices that undercut Western humanoid manufacturers by an order of magnitude. For context, Goldman Sachs estimated the manufacturing cost of a humanoid at $50,000-$250,000 in 2024, a range that has already compressed to $30,000-$150,000. Chinese manufacturers are pushing costs even lower, threatening to commoditize humanoid hardware the same way they did with drones, electric vehicles, and solar panels.

At the consumer end, both Figure AI’s Figure 03 and 1X Technologies’ NEO (at $20,000 pre-order) are targeting home use, suggesting that the addressable market extends well beyond factory floors. ABI Research forecasts a humanoid shipment inflection point of 115,000 units in 2026-2027, driven largely by Chinese volume production and declining component costs.

Barriers, Skepticism, and the Labor Question

Despite the momentum, significant technical and economic barriers remain. Battery life is perhaps the most fundamental constraint: current humanoid robots typically operate for two to four hours on a single charge — Tesla’s Optimus V2 manages roughly two hours of dynamic operation, Unitree’s H1 less than four hours of static use, and Agility’s Digit only about 90 minutes. Achieving a full eight-hour shift without recharging could take a decade or longer, though battery-swapping systems (used by both Digit and Apptronik’s Apollo) offer a near-term workaround for continuous operation. Solid-state batteries, being adopted by Chinese manufacturers like Xpeng and EngineAI, are pushing runtimes beyond four hours and could accelerate progress — TrendForce projects humanoid demand for solid-state batteries may surpass 74 GWh by 2035.

Dexterity — the ability to manipulate small, delicate, or irregularly shaped objects — remains far below human capability. While robots can pick up rigid boxes reliably, handling flexible materials (fabric, cables, plastic bags) and performing fine motor tasks (turning screws, connecting small connectors) is still unreliable. Figure 03’s three-gram-sensitivity fingertip sensors represent meaningful progress, but the gap to human-level manipulation remains wide.

On the regulatory front, the landscape is evolving faster than the draft industry narrative suggests. ISO 10218:2025 has been updated with integrated collaborative robot guidance, ANSI released the updated R15.06-2025 standard with new cybersecurity provisions, and ISO 25785-1 is under development specifically for dynamically stable robots like humanoids — addressing unique risks such as fall hazards when power is cut. These are not yet comprehensive humanoid-specific regulatory frameworks, but the standards infrastructure is materializing.

The labor market implications remain profound and contentious. The International Federation of Robotics, in its position paper on humanoid robots, has taken a cautiously optimistic stance, noting that robotization at manufacturing plants actually correlates with a 150% increase in job postings and a 15% rise in employment at adopting facilities — though it emphasizes that the economic viability of humanoids versus traditional automation remains unproven. The IFR projects that by 2034, more than half of manufacturing operators globally will work alongside robots in some capacity. The ILO has urged a worker-centered approach, recommending that unions focus on skills retraining, optimal redeployment of displaced staff, and frameworks for human-robot coexistence rather than opposing automation outright. For developing nations with large manufacturing labor forces, the stakes are particularly high: if humanoid robots make nearshoring to high-cost countries more economical, the comparative advantage of low-wage manufacturing nations could erode significantly.

Frequently Asked Questions

Sources & Further Reading

• Figure AI Exceeds $1B in Series C Funding at $39B Valuation — Figure AI
• Figure 02 Contributed to the Production of 30,000 Cars at BMW — Figure AI
• Introducing Figure 03 — Figure AI
• Musk Admits No Optimus Robots Are Doing Useful Work at Tesla — Electrek
• The Global Market for Humanoid Robots Could Reach $38 Billion by 2035 — Goldman Sachs
• Humanoid Robot Market: $357 Billion Impact Anticipated by 2040 — Morgan Stanley
• An Electric New Era for Atlas — Boston Dynamics
• Boston Dynamics Beats Tesla to the Humanoid Robot Punch — The Register
• Nvidia Announces Isaac GR00T N1 Open Foundation Model — Nvidia Newsroom
• RT-2: New Model Translates Vision and Language into Action — Google DeepMind
• China’s Unitree Ships More Than 5,500 Humanoid Robots in 2025 — South China Morning Post
• Amazon Deploys Over 1 Million Robots — About Amazon
• Digit Moves Over 100,000 Totes in Commercial Deployment — Agility Robotics
• Apptronik and Mercedes-Benz Enter Commercial Agreement — Apptronik
• Figure AI Dumps OpenAI Deal After Major Breakthrough — Decrypt
• Humanoid Robots: Vision and Reality — International Federation of Robotics
• Humanoid Robots Set to Drive Demand for Solid-State Batteries — TrendForce
• Reality Is Ruining the Humanoid Robot Hype — IEEE Spectrum