The State of Robotics (May 2026)

The shape of the field

Sized by capital raised, the field's shape is stark: humanoids, robotaxis/AVs, and defense autonomy hold most of the disclosed money. Flip the toggle to company count and a different picture appears — a wide, mundane base of industrial arms, warehouse/logistics robots, inspection, drones and field machines dwarfs the frontier bets. Robotics is both: a few heavily-funded moonshots sitting on a broad industrial base.

Where the money is

Disclosed capital by use case. The money concentrates in the hardest, most capital-intensive corners — humanoids, robotaxis/AVs, and surgical robotics — where building a robot is a billion-dollar undertaking.

Sum of disclosed funding per use case ($M). Only ~40% of companies disclose funding, so bars understate true capital.

The funding leaders

The most-capitalized companies in the field — robotaxi/AV operators, humanoid labs, and defense-autonomy firms dominate the top of the table.

Top companies by money raised ($M, log scale), colored by HQ region.

Where capital has flowed — over time

The single total hides the story; the flow over time tells it. Robotics funding ramped from ~$1B/yr (2015) to a $31B peak in 2021, cooled through 2022–23, then came roaring back — 2024–26 are among the biggest years on record, but now overwhelmingly via mega-rounds ($100M+), with early-stage flat. Below: capital deployed per year, by region and by use case.

Annual disclosed capital deployed, reconstructed from 1032 companies’ dated funding rounds (~$175.4B — roughly 90% of all disclosed capital is dated). Mega-rounds (≥$100M) now make up ~80–96% of dollars. 2026 is a partial year. Undated or undisclosed rounds are excluded.

Who’s hiring now

A live momentum signal: open job postings pulled directly from companies’ public hiring boards. The pattern is stark — defense, drones, and autonomous vehicles are hiring hardest, led by Anduril (1,500+ open roles), Waymo, Shield AI, and Zoox. Humanoid and manipulation startups are scaling teams fast but off a smaller base.

Live open roles (distinct, deduped by title) from 155 companies with a public Greenhouse / Lever / Ashby job board — this skews to funded, US, and defense/AV firms and covers only a slice of all companies; it is not a hiring census, and a blank is “unknown,” not zero.

When it was built

A young, fast-forming industry: most of today's robotics companies were founded in the last decade, accelerating into the 2020s as learning-based control and cheap compute lowered the barrier to building a capable robot.

Companies by founding year (where known, 1990–2026).

Where it's being built

A tri-pole — the US, China, and Europe — with a dense second tier in Korea, Japan, and Israel. Hover any country for its company count and disclosed capital.

Unique companies by HQ country. Country-level, not pin-level; some countries unmatched by the map renderer are omitted.

Geopolitics & Supply Chain

The robotics industry has split into three unequal zones of gravity — China, the US, and Europe — with Korea, Japan, and Israel forming a capable second tier. That split runs deeper than trade policy: it reflects where each bloc actually controls the stack, from raw materials to software to end-market access.

China dominates hardware volume. In 2024 China installed roughly 295,000 industrial robots — 54% of the global total — and its operational stock crossed 2 million units, nearly half the world's entire base (IFR World Robotics 2025). In humanoids the lead is sharper still: Chinese firms — led by AgiBot (39% share in 2025) and Unitree (32%) — collectively held close to 80% of global humanoid shipments in 2025, according to Bloomberg. More than 140 Chinese humanoid makers now compete domestically, with per-unit costs roughly one-third to one-half of Western equivalents. The competitive moat is a vertically integrated hardware supply chain built up through the EV industry: motors, reducers, sensors, and battery packs sourced within a tight geographic cluster in the Yangtze River Delta. Inovance Technology alone shipped over 5 million robotic joint servo motors in 2025.

The US leads software, foundation models, and defense autonomy. Boston Dynamics' Atlas (now in production, partnered with Google DeepMind on robot foundation models), Figure AI's pilot with BMW, and Microsoft's Rho-alpha foundation model (January 2026) represent a US advantage that is less about unit count and more about AI architecture. On the defense side, Ghost Robotics' Vision 60 quadruped serves US Air Force and Marine Corps perimeter missions; Defense Secretary Hegseth's July 2025 memo directed all service branches to accelerate fielding of drone and robotic systems, with dedicated robotic units mandated by FY2027.

Europe leads precision industrial arms and AMRs. ABB (Västerås/Zurich, ~$3.2B robotics revenue) is the world's largest industrial robot company by revenue. KUKA (Augsburg, owned by China's Midea since 2016) remains Germany's flagship arm-maker. European companies dominate high-value segments: precision automotive welding, cleanroom electronics, and warehouse AMRs, where ABB and KUKA collectively hold leadership positions. The region's strength is engineering depth and reliability credentials rather than volume.

Second tier: Korea, Japan, Israel. South Korea holds the world's highest robot density — 1,220 robots per 10,000 employees — and installs around 30,000 units annually; companies like Hyundai Robotics and Doosan supply both domestic and export markets. Japan (Fanuc, Yaskawa, Kawasaki) remains the second-largest installation market with ~44,500 units in 2024, and Japanese suppliers still dominate high-precision servo motors and harmonic drives globally. Israel punches well above its size in defense robotics: Elbit Systems acquired autonomous off-road vehicle developer Blue White Robotics in May 2026, and its order backlog hit a record $30.2 billion by Q1 2026; IAI and Rafael are fielding AI-based autonomous logistics and counter-drone systems battle-tested in active conflict.

The supply-chain chokepoint: rare-earth magnets. Neodymium-iron-boron (NdFeB) magnets are inside almost every servo motor and actuator in a robot. China controls the overwhelming majority of global rare-earth processing. In April 2025 Beijing imposed a licensing regime covering seven heavy rare earths — including dysprosium and terbium, critical for high-temperature-stable magnets — on dual-use grounds. Those controls remain fully active as of mid-2026. A typical industrial robot servo motor uses 300–500 g of rare-earth-doped magnets; a humanoid uses multiples of that across its joints. A second tranche of controls announced in October 2025 — which would have required licenses for components made anywhere using Chinese rare-earth technology — was temporarily suspended until November 2026, but the architecture is in place.

What "addressable market" means depends on your flag. A Western vendor selling into North America, Europe, Japan, or Korea faces a relatively open field for industrial and commercial robots, but encounters screening risk (CFIUS, export controls on advanced chips) when sourcing Chinese components or partnering with Chinese firms. A Chinese vendor faces the inverse: dominant cost position at home and in the Global South, but a narrowing aperture in Western markets. The bipartisan American Security Robotics Act, introduced in the US Senate in March 2026 by Cotton and Schumer, would ban federal agencies from buying robots made by firms under foreign-adversary jurisdiction — a direct parallel to the Huawei telecoms playbook. For now it has not passed, but the direction of travel is clear. As the geography chart above shows, company formation is still concentrated in the US and China, with Europe clustered around a smaller number of larger incumbents — a distribution that mirrors these structural advantages.

Sources: IFR World Robotics 2025 report; Bloomberg (Chinese humanoid shipments, Jan 2026); Arnold Magnetic Technologies / IEA / RSIS on rare-earth export controls; TechCrunch, Rest of World on China humanoid market; FDD / Orbital Today on American Security Robotics Act (Mar–Apr 2026); Breaking Defense / Israel Defense on Elbit backlog (2026); Boston Dynamics CES 2026 announcement; DARPA news release (2026); US Defense Secretary Hegseth memo (Jul 2025). Some market-share figures vary by source; ranges are noted where estimates diverge.

Deployment reality — pilots vs. production

The Robot Report's shorthand for 2026 is "less dancing, more doing." The honest picture is a split: a handful of robot categories are genuinely deployed at scale and earning revenue, while humanoids — however much airtime they get — are nearly all still in single-digit or small-fleet pilots. The two halves of that picture are not comparable, and conflating them distorts the field.

Genuinely at scale. Amazon crossed one million warehouse robots in mid-2025 across more than 300 fulfillment centers globally — a fleet that now approaches its 1.5-million-person headcount in raw unit count. These are not humanoids; they are purpose-built mobile drive units, robotic arms (Sparrow), and autonomous floor movers (Proteus) that have been shipping and iterating since 2012. Waymo is running roughly 500,000 paid robotaxi rides per week as of March 2026 across ten U.S. cities, targeting one million weekly rides by year-end; it completed more than 14 million paid trips in 2025 alone. Zipline has surpassed two million commercial drone deliveries across Africa and is actively launching in Houston and Phoenix; a Zipline drone takes off somewhere in the world every 50 seconds. Brain Corp's BrainOS platform powers an estimated 35,000–40,000 autonomous floor-scrubbing robots operating in retail, airports, and logistics facilities. Carbon Robotics' LaserWeeder has treated more than 250,000 acres across 200-plus units deployed on U.S. farms, with a second-generation modular line launched in early 2025. Serve Robotics scaled from roughly 100 to 2,000-plus sidewalk delivery robots in a single year (FY2025); Starship Technologies operates around 2,700 robots across 150-plus locations and has logged nine million-plus deliveries.

Real commercial deployments, not yet at scale. Agility Robotics' Digit humanoid is operating under a commercial Robots-as-a-Service contract at a GXO Logistics facility in Flowery Branch, Georgia, where it has moved more than 100,000 totes — the clearest example of a humanoid in genuine revenue-generating work, though still a single site. Schaeffler has made a strategic investment in Agility and is planning broader rollout across its 100-plant network, but that is in early stages. Figure AI's Figure 02 completed an 11-month deployment in BMW's Spartanburg body shop, contributing to production of more than 30,000 BMW X3s and moving 90,000-plus sheet-metal components; BMW is now running follow-on pilots in Leipzig and evaluating Figure 03, but the Spartanburg project was explicitly described as a test, not a production line changeover.

Still largely in the lab or in pilots. Most U.S. humanoid deployments in 2026 involve one to ten units operating in controlled or semi-controlled conditions. Gartner (January 2026) predicted fewer than 20 companies will achieve genuine production-scale humanoid deployment in manufacturing or supply chain by 2028. Tesla's Optimus, 1X's Neo, Apptronik's Apollo, and most of the Chinese humanoid field (with the partial exception of Unitree and AgiBot shipping units into research and light-industrial settings) are pre-commercial or in early contracted pilots. The bottlenecks are real: dexterous manipulation in unstructured environments, battery runtime, and — most critically — the labor-intensive task of writing robust policies for every new task in every new environment.

What "scale" actually means here. For mobile logistics robots, scale is tens of thousands of units. For autonomous vehicles, it is hundreds of vehicles and hundreds of thousands of rides per week. For humanoids, "scale" in 2026 means tens of units at most sites, with a few outliers. The categories are on entirely different adoption curves, and the gap matters for anyone trying to assess where robotic deployment actually stands.

Sources: Amazon/GeekWire (July 2025); Waymo blog and TechCrunch (March 2026); Zipline newsroom (early 2026); Brain Corp/North America Outlook; Carbon Robotics/BusinessWire (Feb 2025); Serve Robotics FY2025 report; Starship Wikipedia/Robotics & Automation News; Agility Robotics (agilityrobotics.com); BMW Group press release (Spartanburg, Nov 2025; Leipzig 2026); Gartner press release (Jan 2026). Fleet and trip counts are as of the most recently reported figures and may have changed.

Where the intelligence runs

A defining technical fact about robots: the fast control loop almost always runs on the robot (a dropped packet at 200 Hz is a fall or a crash). The heavier "thinking" — planning, perception reasoning — runs onboard, in a hybrid edge/cloud split, or (rarely, for latency-tolerant work) in the cloud. Training almost always happens in the cloud. This chart shows where each robot type's thinking layer sits today.

Bubble = a use case; x = where its thinking layer runs; size = # companies; color = placement (red onboard · amber hybrid · green cloud).

How robots learn — a primer

Modern robots are less hand-coded and more trained, the way a chatbot is trained. Instead of an engineer writing rules for every situation, a robot watches examples or practices a task and learns a "policy" — a function that turns what it sees into what it does. This section explains the handful of ideas you'll meet repeatedly in the rest of the report.

Three places the AI runs. It helps to picture three layers, each working at a different speed. (1) Fast control on the robot — a tight loop running hundreds to a thousand times a second that keeps the machine balanced and its joints tracking targets; this is reflex-speed and runs on a chip onboard. (2) The "thinking layer" — a slower model, often running a few to a few dozen times a second, that looks at the cameras, reads the instruction ("put the cup in the sink"), and decides what to do next. (3) Training in the cloud — the heavy, offline work of learning from large datasets on big GPU clusters; the result is shipped down to the robot. Figure's Helix (Feb 2025) made this split explicit with a "System 2" reasoning model at ~7-9 Hz feeding a "System 1" motor controller at 200 Hz; NVIDIA's GR00T N1 (Mar 2025) adopted the same two-system design.

Vision-Language-Action (VLA) models. The headline idea of the current era. A VLA takes in camera images plus a text instruction and directly outputs robot actions — it fuses "what a vision-language model knows about the world" with "how to move." Notable examples: RT-2 (Google, 2023), the open-source OpenVLA (2024, 7B parameters), and π0 ("pi-zero," Physical Intelligence, late 2024), with an upgraded π0.5 released in 2025. The pitch is generalization: because the model inherits web-scale knowledge from its language backbone, it can attempt objects and instructions it never saw in robot data.

Imitation learning vs. reinforcement learning. These are the two main ways a policy is taught. In imitation learning (also "behavioral cloning"), the robot learns by copying human demonstrations — you show it the task many times and it learns to reproduce the motions. It is the dominant method for manipulation today because it's data-efficient and stable, but the robot is only as good as the demos. In reinforcement learning (RL), the robot practices by trial and error, getting a reward signal for success; it can discover better-than-human strategies but needs huge numbers of attempts, so it's mostly done in simulation (common for legged locomotion and balance).

Teleoperation is how most imitation-learning data gets made: a person remotely puppeteers the robot — moving its arms or wearing a control rig — while every camera frame and joint position is recorded. Each run is one "demonstration." Stanford's low-cost ALOHA and Mobile ALOHA rigs popularized this, and the open-source LeRobot library (Hugging Face) standardized the collect-train-deploy loop on cheap hardware.

Action chunking is a trick that made imitation learning work far better. Rather than predicting one tiny next move at a time (which lets small errors snowball), the model predicts a short chunk of future actions at once — typically dozens of steps. Introduced with ACT (Action Chunking with Transformers, the ALOHA paper) and now used almost everywhere, including in π0 and OpenVLA.

Diffusion policy & flow matching. Two closely related techniques for generating those action chunks smoothly. Diffusion policy borrows the math behind AI image generators: it starts with random noise and "denoises" it step by step into a clean sequence of motions, which is good at handling tasks that have several valid ways to do them. Flow matching is a newer, faster cousin that learns a more direct path from noise to action; π0 uses it to output smooth trajectories at 50 Hz.

Sim-to-real means training in a physics simulator — where you can run millions of cheap, safe attempts — and then transferring the learned skill to a physical robot. The catch is the "reality gap": simulators are never quite right, so skills can break in the real world. Techniques like randomizing the simulation and, increasingly, generative world models aim to close that gap. NVIDIA's Cosmos world foundation models (announced CES 2025, updated through 2025) and the Isaac Sim/Omniverse stack are the most visible examples.

World models are AI systems that learn to predict what happens next in the physical world — essentially a learned simulator or "imagination." A robot (or a training pipeline) can roll the world forward in its head to plan, generate synthetic training data, or test a plan before acting. This is an active frontier rather than a solved tool.

"End-to-end" and "foundation model," for robots. End-to-end means a single learned model goes straight from raw sensors (pixels) to motor commands, with no hand-built pipeline of separate perception, planning, and control modules in between. A foundation model for robots is one large model pretrained on broad, diverse data (many tasks, many robot types) that can then be fine-tuned for specific jobs — the same idea as a foundation model in language, applied to physical action. VLAs like π0 and GR00T are the leading attempts at robot foundation models.

VLA (Vision-Language-Action model)

A single model that takes camera images plus a text instruction and outputs robot actions. Examples: RT-2, OpenVLA, π0.

Imitation learning

Teaching a robot by having it copy recorded human demonstrations (also called behavioral cloning). The dominant method for manipulation today.

Reinforcement learning (RL)

Teaching a robot through trial and error guided by a reward signal. Powerful but data-hungry; mostly done in simulation, common for locomotion.

Teleoperation

A human remotely controlling the robot to record demonstration data. The main source of imitation-learning datasets (e.g., ALOHA, LeRobot).

Action chunking

Predicting a short sequence of future actions at once instead of one step at a time, which reduces error buildup. Introduced with ACT.

Diffusion policy

Generating an action sequence by starting from noise and denoising it, borrowing the technique behind AI image generators; handles tasks with multiple valid solutions.

Flow matching

A faster relative of diffusion that learns a more direct path from noise to actions; used by π0 for smooth real-time control.

Sim-to-real

Training in a physics simulator and transferring the skill to a real robot; the difficulty is the "reality gap" between sim and the physical world.

World model

A learned model that predicts what happens next in the physical world — a kind of imagination used for planning or generating training data (e.g., NVIDIA Cosmos).

Foundation model (for robots)

A large model pretrained on broad, diverse robot data that can be fine-tuned for specific tasks — the robotics analog of a language foundation model.

End-to-end

A single learned model that maps raw sensor input directly to motor commands, with no hand-built modules in between.

Concepts and examples verified via project/model pages and papers (May-Jun 2026): Physical Intelligence (π0/π0.5), Figure (Helix), NVIDIA (GR00T N1, Cosmos, Isaac Sim), OpenVLA, RT-2, ALOHA/ACT, and Hugging Face LeRobot. Hz figures and parameter counts are as stated by their sources; some are approximate.

The field's lineage — a timeline

Modern robot learning moved fast, and most of the ideas behind today's systems landed in a tight four-year window. The throughline: borrow the recipe that worked for language and vision models — pretrain a big model on lots of data, then adapt it — and apply it to actions. The result is the "vision-language-action" (VLA) model: feed in camera images plus a plain-language instruction, get out motor commands. The table below traces the milestones that got the field here. Dates are the public release or paper date.

Read top to bottom, the pattern is clear: 2022–2023 proved big models could learn robot actions; 2024 made the data and base models open; 2025 turned them into shipped foundation models and first real deployments; and the latest shift is adding reinforcement learning so robots keep improving after the initial training.

Sources: papers/arXiv and official posts for each entry — Google DeepMind (RT-1, RT-2, Gemini Robotics, "Self-Improving Embodied Foundation Models"), RSS 2023 (Diffusion Policy), Stanford (ALOHA/ACT, Mobile ALOHA), Open X-Embodiment (arXiv 2310.08864), Octo & OpenVLA (arXiv 2406.09246), Physical Intelligence (π0 arXiv 2410.24164, π0.5 arXiv 2504.16054, openpi, π*0.6/RECAP), NVIDIA (Project GR00T, Cosmos, Isaac GR00T N1/N1.5 newsroom), Hugging Face (SmolVLA arXiv 2506.01844), Agility Robotics/GXO and Figure AI/BMW deployment posts. Dates are public paper/release/announcement dates; "production pilots" are pilots, not permanent fleet installs. The 2025–2026 RL-tooling row reflects an ongoing trend, not a single product launch.

Where the research is heading

The three flagship 2026 reports and an open-source corroboration pass agree on roughly the same live frontier. The headline is that the hard problem has moved from the robot's body to its brain — specifically, from building capable hardware (now widely available and increasingly cheap) to teaching one piece of software to do many tasks, on many robots, reliably enough to deploy. A few directions recur across every source:

Vision-Language-Action (VLA) models are now the center of gravity. Instead of hand-coding each task, teams take a model that already understands images and language, and add a third channel — robot actions — so it can be told what to do in plain words and then move accordingly. This shift, which crossed from research into real deployments in 2025–2026, is what most of the field's other work now orbits.

Data, not algorithms, is the binding constraint. Robots learn mostly by copying human demonstrations (imitation learning) collected through teleoperation — a person remotely puppeteering the robot. The cost of that data has fallen sharply (reported around $340/hr in 2024 to roughly $118/hr in early 2026), and imitation learning has overtaken reinforcement learning as the main way manipulation skills are taught. To stretch scarce real data further, the field pools demonstrations across many robot types (so one policy transfers between different machines) and increasingly manufactures synthetic data.

World models are emerging as the synthetic-data engine. These are learned simulators that can imagine realistic video and physics, letting teams generate huge volumes of training trajectories cheaply and letting a robot "rehearse" an action before doing it. Reports expect them to become a standard ingredient by 2027.

Architectures are converging on a fast/slow split. A slower, high-level model "thinks before acting" — reasoning about the goal and breaking it into steps — and hands sub-goals to a fast, low-level controller that produces smooth motion in real time. For generating that motion, flow matching (a newer technique) is displacing plain diffusion because it yields smoother continuous control.

Bigger isn't automatically better. Unlike large language models, robot foundation models reportedly hit diminishing returns past roughly 7B parameters, so a strong push toward efficiency — compact models that run on the robot itself, plus tricks to keep them acting while the next move is computed — runs alongside the push for capability.

The named open frontiers the reports still flag as unsolved: dexterous manipulation (fine, contact-rich hand skills), whole-body humanoid control, long-horizon and compositional tasks (stitching many skills together and recovering from mistakes without reprogramming), generalization and robustness in the messy real world, an emerging reinforcement-learning post-training stage layered on top of imitation (mirroring how LLMs are fine-tuned), and — repeatedly stressed — evaluation, safety and standards, where bodies like NIST are only now beginning to build the "Consumer Reports for humanoids" that the field lacks.

Synthesized from the roboticscenter.ai, F-Prime Capital, and Robot Report 2026 outlooks plus an open-source-model survey (May 2026).

The open-source stack — models & techniques

One striking pattern across the sources: the strongest robot "brains" are increasingly open-weight — published with downloadable model weights anyone can fine-tune — closing the gap with closed, proprietary systems. Below are the open-source models named most consistently across the reports, followed by the handful of techniques the field has converged on. (A model's parameter count is a rough proxy for size; smaller can run on cheaper hardware.)

The dominant techniques (what these models actually do under the hood):

• Vision-Language-Action (VLA) policies — a vision-language model adapted to output robot actions; the central paradigm.
• Imitation learning / behavior cloning from teleoperation — learning by copying human demonstrations; the workhorse data loop (open hardware like ALOHA and SO-100/101).
• Action chunking — predicting a short sequence of future moves at once rather than one step, which smooths motion and cuts error build-up (originated with ACT).
• Diffusion policy & flow matching — two ways of generating smooth, multi-modal continuous motion; flow matching is now the preferred choice in frontier models.
• Cross-embodiment training — pooling data across many robot types so one policy transfers between different machines.
• Dual-system "think-before-act" — a slow high-level planner paired with a fast low-level controller for real-time control.
• Sim-to-real transfer — training in simulation (with randomized physics/visuals) then deploying on real hardware.
• World-model / synthetic-data generation — using learned simulators to amplify scarce real data (reportedly ~40% gains when mixed with real data).

Open-weight models named most consistently across the 2026 reports; each verified current as of May 2026.

The data bottleneck

Algorithms for robot learning have matured rapidly; the practical ceiling is now the scarcity and cost of high-quality training demonstrations. A language model can train on the full text of the internet. A manipulation policy needs data of a robot doing the task — and that data barely exists at scale outside the handful of labs and companies that have spent years collecting it.

Why teleoperation dominates collection. The dominant method is still direct teleoperation: a human operator controls the robot through a leader arm or VR interface while the robot's sensors record everything. The data is dense, labelled by construction, and covers the physical nuances that simulation cannot yet replicate. The problem is cost. Industry estimates (Silicon Valley Robotics Center, 2026) put the fully loaded price of one hour of high-quality teleoperation data at roughly $340 in early 2024. By March 2026 that figure had fallen to around $118 per hour, a 65% drop driven mainly by cheaper hardware and more efficient annotation pipelines. That is still expensive: most manipulation tasks need 300–1,200 demonstrations, so a single-task dataset can run $35K–$150K before you touch model training.

Hardware that made it cheaper. The cost curve bent when researchers started publishing open hardware designs. Stanford's ALOHA system (2023) and its bimanual mobile successor Mobile ALOHA demonstrated that a capable leader-follower rig could be built for roughly $20K–$32K rather than the six-figure exoskeleton setups common in industrial labs. The SO-100 and SO-101 arms — designed by TheRobotStudio and integrated into Hugging Face's open-source LeRobot framework — pushed further, making a functional leader-follower pair accessible for a few hundred dollars in parts. These open platforms have seeded a community of small labs and hobbyists generating demonstration data that would not otherwise exist.

Open datasets: progress and limits. The largest public corpus is Open X-Embodiment (Google DeepMind and 21 institutions, arXiv October 2023): over 1 million real-robot trajectories drawn from 60 existing datasets, spanning 22 robot embodiments and 527 demonstrated skills. It proved that cross-embodiment training is feasible and gave the field a shared baseline. DROID (Stanford, published March 2024, presented at RSS July 2024) took a different approach — 76,000 trajectories, 350 hours, 564 scenes, collected in-the-wild across three continents by 50 operators, all on the same Franka Panda hardware. DROID's diversity of environments rather than diversity of robots is its distinguishing feature. Both datasets are openly available, and both immediately exposed how thin public coverage remains: 1 million episodes sounds large until you compare it to the billions of tokens in a language model pretraining corpus.

Supplementary sources. Three other approaches are in active use. Fleet logs — data from robots already deployed commercially — are the most valuable because they are real, varied, and generated at low marginal cost once hardware is in the field. Internet video is being mined for priors about how objects behave and how humans move; models like RT-2 showed that vision-language pretraining on web data transfers to robot control, though the embodiment gap limits how directly video translates to action. Synthetic and world-model data is the fastest-growing frontier: NVIDIA's Cosmos platform and a range of physics simulators generate sensor-realistic trajectories at scale, and video-generation world models are beginning to serve as learned simulators for policy rollout. None of these fully substitute for real contact data yet, but they reduce how much of it is needed.

The moat question. Open datasets let anyone get started. They do not level the playing field for deployment. A company running robots in a specific environment — a particular warehouse layout, a particular product mix — accumulates demonstrations that are precisely tuned to that context. That vertical dataset improves the policy, and a better policy runs more hours, generating more data. The compounding effect is why proprietary data pipelines are consistently cited as the primary defensibility argument in the sector: algorithms can be replicated; a year of contact-rich demonstrations from real operations cannot.

Teleoperation cost figures (~$340/hr 2024, ~$118/hr March 2026) are industry estimates from the Silicon Valley Robotics Center State of Robotics 2026 report, not peer-reviewed; treat as directional. Open X-Embodiment statistics from robotics-transformer-x.github.io and arXiv:2310.08864 (October 2023). DROID statistics from droid-dataset.github.io and arXiv:2403.12945 (March 2024). ALOHA from arXiv:2304.13705 and arXiv:2401.02117. SO-101 from github.com/TheRobotStudio/SO-ARM100 and Hugging Face LeRobot.

Benchmarks & simulation — how progress is measured

Robot "progress" is hard to pin down because there is no single agreed-upon scoreboard. In practice, capability is measured two ways: success rate on standardized tasks (did the arm actually pour the cup, fold the cloth, or complete the multi-step chore?) and, increasingly, head-to-head comparisons between policies. A lot of this evaluation happens in simulation — virtual physics worlds where a controller can attempt a task thousands of times overnight — because running the same trials on real hardware is slow, expensive, and easy to set up inconsistently from one lab to the next.

Benchmarks are shared task suites with fixed rules, so different research groups can report numbers that are at least roughly comparable. Most are simulated and focus on tabletop manipulation (picking, placing, opening drawers), though newer ones push toward long, multi-step household chores and toward grading models on real robots.

Simulators are the physics engines and environments where policies are trained and stress-tested. The main ones in use today:

• NVIDIA Isaac Sim / Isaac Lab — GPU-accelerated, photorealistic simulation widely used for training and synthetic data; Isaac Sim 5.0 and Isaac Lab 2.2 became generally available in 2025.
• MuJoCo — a fast, accurate physics engine (open-sourced by Google DeepMind in 2022); its MJX variant runs many simulations in parallel on GPUs/TPUs. Underpins many other tools.
• SAPIEN / ManiSkill — an open GPU-parallelized manipulation simulator and benchmark; ManiSkill3 (2025) emphasizes very high-throughput rendering for visual learning.
• Genesis — an open-source engine first released December 2024, notable for very fast physics and generating scenes from natural-language prompts.
• NVIDIA Cosmos — not a classical simulator but a family of "world foundation models" that generate physics-aware video to expand training data; early adopters include Figure and Agility Robotics.

The honest caveat: there is still no trusted, standardized "Consumer Reports for humanoids" — no neutral body-of-record that rates whole robots head-to-head the way crash tests rate cars. Sim scores don't always carry over to real hardware (the "sim-to-real gap"), and companies often report their own numbers on their own tasks. Efforts to fix this are underway: the U.S. National Institute of Standards and Technology (NIST) proposed a baseline performance benchmark for humanoid robots in May 2026 — a set of locomotion and manipulation tasks with quantifiable metrics — but it is still in the proposal stage and seeking consensus, not yet a finished standard.

Sources: arXiv papers and project pages for LIBERO, Meta-World, BEHAVIOR-1K, CALVIN, SimplerEnv, ManiSkill3, RoboArena; NVIDIA developer blog (Isaac Sim 5.0 / Isaac Lab 2.2, Cosmos); Google DeepMind MuJoCo; Genesis project; NIST and The Robot Report (humanoid baseline benchmark proposal, May 2026). Verified May–Jun 2026; NIST benchmark is a proposal, not a finalized standard.

Safety, standards & governance

Robots that share space with people have to prove they won't hurt anyone, but the rulebook was written for an earlier generation of machines. Today's developers borrow safety standards from adjacent categories of robots and stretch them to fit. There is still no humanoid-specific safety standard in force — that gap is the defining feature of the current landscape.

The borrowed standards. The functional-safety framework robots lean on is a stack of ISO and ANSI documents, each written for a particular kind of machine:

ISO 10218 (industrial robots)

The bedrock standard for factory robot arms. A major revision landed in 2025 (ISO 10218-1:2025 for the robot itself, ISO 10218-2:2025 for the cell and integration) after roughly eight years of work by experts from 20+ countries. Notably, the 2025 edition folded in collaborative-operation requirements and dropped the separate "collaborative robot" category.

ISO/TS 15066 (collaborative robots)

A 2016 technical specification that set the force and pressure limits a "cobot" may apply to specific parts of the human body before it must stop — the reference data behind power-and-force-limiting designs. As of the 2025 revision, its requirements are being absorbed into ISO 10218-2 rather than living as a standalone document.

ISO 13482 (personal-care robots)

The 2014 standard for robots that physically assist people — mobility aids, lift assists, exoskeletons. It is the closest existing standard to a "robot living among people," but it is limited to earthbound personal-care use and predates modern humanoids. A revision (toward broader "service robot" scope) is in development.

R15.08 / ANSI (industrial mobile robots)

The U.S. standard for autonomous mobile robots (AMRs) and mobile manipulators, published in parts: ANSI/RIA R15.08-1-2020 covers the robot, R15.08-2-2023 covers systems and integration, with a third part (for users) planned. It exists because older AGV rules assumed fixed guide paths, not free navigation.

The humanoid gap. None of the above was designed for a free-walking, two-legged robot in a warehouse or a home. Bipeds are "dynamically stable" — they need active control just to stay upright — which makes proving their safety far harder than for a bolted-down arm. Standards bodies are organizing, but slowly. ISO's robotics committee (TC 299) is drafting ISO 25785-1, aimed at industrial mobile robots with actively controlled stability (legged or wheeled); per ISO's own catalog it remains at working-draft stage, and it covers industrial use, not consumer ownership. Separately, an IEEE study group (chaired by ASTM International's Aaron Prather, 60+ contributors) published a framework report in September 2025 on classification, stability metrics, and human-robot interaction — explicitly "a starting line, not a finish line," with ratified standards estimated to be 18–36 months out. For now, makers self-certify against the borrowed standards above.

Liability and insurance. Because product-liability allocation for autonomous robots is unsettled, insurers and integrators are improvising. As robots connect to enterprise networks and cloud services, a technical fault can become a business and legal event — downtime, regulatory questions, contract disputes across the supply chain. The EU's Machinery Regulation (2023/1230, applying from 2027) adds mandatory cybersecurity requirements where safety functions are affected.

Data privacy. A humanoid or service robot is a roaming sensor platform — cameras, microphones, depth sensors, sometimes always-on. In the EU, the AI Act treats many autonomous robots as high-risk AI systems, triggering data-governance and human-oversight duties. In the U.S., obligations are arriving piecemeal at the state level — California's frontier-AI transparency law (effective Jan 1, 2026), Colorado's high-risk-AI law (effective mid-2026), and Texas's TRAIGA (signed 2025) — rather than through one federal rule.

Defense and export. Robotics sits on the ITAR/EAR line: ITAR governs military-specific items and technical data, while the EAR covers dual-use technology with both civil and military applications — where most commercial robots and their software fall. The line moves: a September 2025 ITAR final rule reworked U.S. Munitions List entries and added a license exemption for certain unmanned underwater vehicles, a reminder that autonomy and export control are actively being renegotiated.

Sources: ISO.org catalog entries for ISO 10218-1/-2:2025, ISO/TS 15066:2016, ISO 13482:2014, and ISO/WD 25785-1; ANSI/A3 R15.08-1-2020 and R15.08-2-2023; The Robot Report and IEEE/ASTM coverage of the Sept 2025 IEEE humanoid framework; EU Machinery Regulation 2023/1230 and EU AI Act; U.S. state AI laws (CA, CO, TX); DDTC ITAR final rule (Sept 2025). ISO 25785-1 status per ISO.org is working-draft; some secondary outlets describe it differently, so treat its publication timing as unsettled.

Company directory — every company, by use case

The full census — every company we mapped, grouped by use case then by region, with what it does, HQ country, disclosed funding, stage and founding year. ~3,074 listings; a company serving several use cases is listed under each (~2,522 unique firms — the charts above are deduped to unique). — = undisclosed.