The State of Embodied Intelligence: Robotics in 2026

The Inflection Point

Embodied AI, the project of giving machines the ability to perceive, reason about, and physically interact with the real world, reached an inflection point in 2025-2026. The global market hit $4.44 billion in 2025 and is growing at 39% annually, projected to reach $23 billion by 2030. Funding rounds exceeded $6 billion in just seven months. NVIDIA declared a "ChatGPT moment for robotics" at CES 2026.

But there is a significant gap between the headlines and the hardware. Most humanoid robots last 90 minutes on a charge. Policies that work 95% of the time in the lab drop to 60% in the real world. Even the companies building humanoid robots told the Wall Street Journal in December 2025 that they think they are overhyped.

This report examines where embodied intelligence actually stands: what works, what does not, and what the $6 billion is buying. It draws on 138 sourced claims across five research domains: foundation models, hardware platforms, commercial deployments, research challenges, and industry dynamics.

What the Evidence Shows

1. VLA models are the new dominant paradigm for robot intelligence Vision-Language-Action models exploded in 2025-2026. ICLR 2026 received 164 VLA paper submissions, an 18x increase from just 9 the year before. NVIDIA, Physical Intelligence, Google, and Ant Group all released major VLA models. But current production models remain small (2 to 7 billion parameters) and the gap between open-source research VLAs and proprietary frontier models in real-world generalization is significant.
2. Funding is consolidating into fewer, larger bets Robotics startups raised over $6 billion in the first seven months of 2025, exceeding the full-year 2024 total. But deal count dropped 30%, from 671 rounds in 2023 to 473 in 2024. The money is concentrating: Figure AI ($1B at $39B valuation), Physical Intelligence ($600M at $5.6B), SoftBank's $5.4B acquisition of ABB Robotics.
3. Specific verticals have achieved real deployment scale Amazon operates over 1 million warehouse robots. Surgical robotics reached 60% adoption in large hospitals. Starship completed 8 million autonomous deliveries. These are not pilot programs; they are production systems.
4. Humanoid robots are in early pilots, not production Tesla has 1,000+ Optimus units in its own factories. Figure AI's 11-month BMW trial contributed to 30,000 vehicles before Figure exited with no production commitment. All public humanoid demos remain teleoperated. The pilot-to-production gap remains wide.
5. Battery life is the critical hardware bottleneck Current humanoids run 90-120 minutes per charge. Industrial use cases need 8-20 hours. Solid-state batteries at meaningful scale are projected for 2035. No near-term solution exists.
6. The sim-to-real gap causes severe reliability drops A robot policy that succeeds 95% of the time in a lab drops to 60% in deployment due to differences in lighting, textures, and camera angles. Production environments require 99.9% reliability. At 95% accuracy, a warehouse robot fails roughly 50 times per day.
7. China deploys 5x more factory robots than the US China had 2 million factory robots in operation in 2024 versus 394,000 in the United States. Chinese company AgiBot led global humanoid shipments at 5,100 units. Meanwhile, 90% of key robotics components still originate from China.
8. Benchmark reproducibility is a credibility crisis NVIDIA's own robotics leadership stated that researchers "pick the best-looking demo from 100 failures." Success rates of 1% are common. Researchers choose benchmark subsets where their models perform best. The field lacks credible, standardized progress metrics.

The Software Revolution: VLA Models

The robotics software stack underwent a paradigm shift in 2025-2026. Vision-Language-Action architectures (models that take in visual input and language instructions, then output physical actions) emerged as the dominant approach, following the same trajectory that large language models blazed in natural language processing.

NVIDIA's January 2026 CES announcement of Isaac GR00T N1.6 set the tone: a 32-layer diffusion transformer trained on thousands of hours of teleoperation data across multiple robot bodies. CEO Jensen Huang called it "the ChatGPT moment for robotics." Physical Intelligence released pi-0.5 with meaningful open-world generalization. Ant Group trained LingBot-VLA on 20,000 hours of bimanual manipulation data.

The platform ecosystem is consolidating around a few players. NVIDIA aims to become the "Android of robotics" by integrating its Isaac and GR00T technologies into Hugging Face's LeRobot framework, connecting 2 million NVIDIA robotics developers with 13 million Hugging Face AI builders. The Robot Operating System (ROS 2) reached 65% market share, up from 40% in 2022, with the global ROS market valued at $572 million.

Open-source models show promise. OpenVLA, trained on 970,000 trajectories, outperforms Google's RT-2-X (55 billion parameters) by 16.5% in task success rate despite being 7x smaller. But a critical gap exists: open-weight research VLAs match proprietary models on simulation benchmarks, yet fall far short in real-world, zero-shot generalization. Only models from Physical Intelligence dominate leaderboards for unseen tasks. Simulation benchmarks may not capture what matters in the real world.

Current VLA Model Landscape

Model	Org	Parameters	Training Data
GR00T N1.6	NVIDIA	2.2B	Thousands of hours, multi-embodiment
pi-0.5	Physical Intelligence	3B	7 platforms, 68 tasks, 104 homes
OpenVLA	Open-source	7B	970K trajectories
GEN-0	Generalist AI	1-10B+	270,000+ hours, growing 10K hrs/week
LingBot-VLA	Ant Group	--	20,000 hours bimanual data

The Hardware Reality

Compute: Edge AI Gets Serious

Running large VLA models on a robot requires serious edge compute. Current models need 50-100 milliseconds per inference, enabling only 10-20Hz control frequency, but manipulation tasks need 20-100Hz. The hardware is catching up: NVIDIA's Jetson T4000, announced at CES 2026, delivers 1,200 TFLOPS within a 40-70W power envelope at $1,999 per unit, a 4x performance gain over the previous generation. Qualcomm entered the humanoid market with the Dragonwing IQ10 featuring an 18-core CPU.

Actuators: 60-70% of the Cost

Actuators (the motors and gearboxes that make robots move) account for 60-70% of a humanoid robot's total manufacturing cost. Specialized vendors like RealMan (200 Nm/kg torque density) and CubeMars (121 Nm/kg, $149.90 per unit) are driving down prices. Traditional manufacturers are entering the market: LG, which makes 41 million motors annually for appliances, targets an actuator product launch in 2027. Hyundai Mobis actuators are already installed in Boston Dynamics' Atlas.

But efficiency remains poor. Electric motors paired with gearboxes achieve 80% motor efficiency, which drops to 40% once the gearbox is included. More than half the energy in a robot's actuation chain is lost as heat.

Sensors: Scaling Production

Sensor manufacturing is scaling fast. Hesai announced plans to double LiDAR production from 2 million to over 4 million units annually, with fully automated lines running a 10-second cycle time per unit. A new Bangkok factory will diversify production outside China by early 2027. LG Innotek is delivering integrated sensing modules combining cameras, LiDAR, radar, and software, with early revenue in the "hundreds of billions of won."

Tactile sensing, while still rare in production, shows dramatic benefits where implemented: 21.9% improvement on pick-and-place tasks, and 90% success on insertion tasks versus 25-40% for vision-only approaches.

Supply Chain Dependency

Roughly 90% of key robotics components still originate from China as of 2026, despite geopolitical pressures. FANUC opened a $110 million Michigan campus and Yaskawa Motoman completed a $200 million Ohio expansion, but nearshoring is a multi-year effort. The US plans to ban Chinese autonomous vehicle software in 2027 and hardware in 2030, creating significant uncertainty for the robotics supply chain.

What Is Actually Deployed

The gap between what is deployed at scale and what makes headlines is vast. The most successful robotics deployments are purpose-built systems in specific verticals, not general-purpose humanoids.

Warehouse Automation: The Clear Leader

Amazon's fleet crossed 1 million robots in June 2026, assisting 75% of global deliveries. Their Sequoia system identifies and stores inventory 75% faster. The DeepFleet AI model improved fleet travel time by 10%. Specialized robots handle specific tasks: Hercules lifts 1,250 pounds, Pegasus uses precision conveyors for packages, and Proteus navigates autonomously alongside employees.

Surgical Robotics: Mainstream Adoption

Over 60% of large hospitals worldwide have integrated surgical robotics. Robotic-assisted procedures represent 55% of complex surgeries in developed countries. Intuitive Surgical dominates with 8,000+ da Vinci units and 12 million procedures. Medtronic's Hugo system secured FDA clearance in urology in December 2025, becoming the first serious US challenger. The surgical robotics market is projected to reach $14 billion by 2026.

Autonomous Delivery: 8 Million and Counting

Starship Technologies completed over 8 million autonomous deliveries with 2,000+ robots across 150+ locations in six countries, crossing 125,000 roads daily. They partnered with Uber Eats for UK delivery. This represents the most mature small-robot outdoor deployment by a wide margin.

Humanoid Pilots: Promising but Early

The humanoid picture is more nuanced. Tesla deployed 1,000+ Optimus units in its own factories for parts processing, targeting 50,000 by year's end at $20,000-$30,000 per unit. Boston Dynamics committed all 2026 Atlas production to Hyundai and Google DeepMind, and is building a factory capable of 30,000 units per year. AgiBot shipped 5,100 humanoids in 2025 with 39% global market share. Global installations reached 16,000 units.

Retail: Clear ROI

Simbe's Tally robot shows concrete returns: stores report a 2% sales increase, 98% on-shelf availability, 50 hours saved per week, and a 20% reduction in out-of-stocks on high-margin items. Agricultural robotics adoption is expected to exceed 65% in developed regions by 2026, though harvesting robots face a basic economics problem: if a quarter-million dollar machine is only as fast as two people, the math does not work.

The Reliability Problem

The defining challenge for embodied AI is not capability in the lab. It is reliability in the real world.

A policy that achieves 95% success under controlled conditions drops to 60% when deployed, because the real world has different lighting, backgrounds, object textures, and camera angles. Production environments require 99.9% reliability. The arithmetic is unforgiving: at 95% per-step accuracy, a 10-step manipulation chain succeeds only 60% of the time. A warehouse robot at 95% accuracy fails roughly 50 times per day, requiring constant human intervention that defeats the purpose of automation.

Long-Horizon Task Coherence

The latest models can operate autonomously for approximately 30 minutes before task coherence degrades. AI task duration is doubling every seven months, from one-hour tasks in early 2025 to a projected eight hours by late 2026. But no robot system is expected to reliably execute chains of 10 or more distinct subtasks in unstructured environments without human intervention by the end of 2026.

The Dexterity Gap

Human hands have 27 degrees of freedom. Achieving human-like functionality requires at least 19-23 in a robotic hand. Robotics pioneer Rodney Brooks observes that "we have not actually seen any improvement in widely deployed robotic hands or end effectors in the last 40 years" despite impressive lab demonstrations. He predicts deployable dexterity will remain "pathetic compared to human hands" beyond 2036.

Sim-to-Real: Getting Closer, Not There

Simulation-to-real transfer made progress. NVIDIA developed GR00T N1.5 in 36 hours using synthetic training data, compared to three months for manual collection. DreamZero achieved 42% relative improvement with only 10-20 minutes of video demonstrations. But these remain exceptions. The persistent failure modes (friction parameters, contact physics, sensor noise) make sim-to-real a problem of engineering detail, not algorithmic breakthrough.

The Inference Bottleneck

Large VLA models require 50-100 milliseconds per inference on edge hardware, enabling only 10-20Hz control. Manipulation tasks typically need 20-100Hz. Until edge compute catches up or models get more efficient, there is a fundamental mismatch between what the AI can think and how fast the robot needs to move.

The Money and the Map

Funding Concentration

The numbers tell a story of consolidation. Over $6 billion raised in seven months, but in fewer rounds. The largest bets:

Company	Round	Amount	Valuation
Figure AI	Series C (Sep 2025)	$1B	$39B
Physical Intelligence	Series B (Nov 2025)	$600M	$5.6B
Apptronik	Series A (Mar 2025)	$403M	--
SoftBank / ABB	Acquisition (Oct 2025)	$5.375B	--
Mobileye / Mentee	Acquisition (Jan 2026)	$900M	--

Industrial customers are becoming investors to secure supply: Mercedes-Benz, Japan Post Capital, and Hyundai have all taken stakes in humanoid companies. This is a signal that strategic buyers believe the technology is close enough to matter for supply chain planning, if not close enough to deploy today.

Geographic Split

The global landscape is split. China dominates in deployment: 2 million factory robots versus 394,000 in the US, a 5x gap. AgiBot leads humanoid shipments globally. But North America captures 37% of embodied AI market revenue, indicating a segmentation between where robots are deployed (China) and where value is captured (US).

The average US robotics engineer earns $133,130, with regional peaks in DC ($147K), California ($147K), and Massachusetts ($145K). European defense robotics investment exceeded $2 billion in 2025, adding a new funding dimension driven by geopolitical pressures.

Key Milestones

How This Research Was Conducted

Voxos Scholar independently investigated five research domains (foundation models, hardware platforms, commercial deployments, research challenges, and industry dynamics) executing 60 search queries and producing 138 unique claims from 140+ source URLs.

Claims were graded by confidence: HIGH (2+ independent sources across scribes, or from official/primary sources), MEDIUM (one credible source, not contradicted), or LOW (single low-authority source, or contradicted by other findings). Cross-scribe validation checked each claim for corroboration across research domains.

Key limitations include heavy reliance on company announcements for 2026 developments, limited access to proprietary deployment metrics, and the benchmark reproducibility issues acknowledged in the research itself, which create uncertainty around claimed performance figures.

Research Sources

Voxos Scholar analyzed 138 claims from 140+ unique sources across 5 research domains: foundation models, hardware platforms, commercial deployments, technical challenges, and industry dynamics. Research conducted February 9, 2026.