Measuring Progress in Robotics and Physical AI

Companion Excel Workbook: Robotics & Physical AI

Measuring Physical AI Progress: Early Milestones and Task Benchmark Results

There is no standard scoreboard for humanoid robots. Software AI has benchmark suites such as MMLU and GPQA and independent aggregators that test every major model the week it ships. Nothing comparable exists for physical AI: testing a robot requires possessing the hardware, a lab, and weeks of work per platform, and most leading humanoids are not for sale. The result is that after four years and tens of billions of dollars of investment, the first independent measurement efforts all appeared within a six-week window in May and June 2026. The field today resembles LLM benchmarking circa 2020, before standard leaderboards existed: capability claims are press releases and demonstration videos, and apples-to-apples comparison across robots is not yet possible.

The individual measured results that do exist are presented below as standalone data points, each with its own source and test conditions. They should not be combined into a single comparison.

Walking speed, independently measured (Fraunhofer IPA, May 2026) Germany’s Fraunhofer IPA published the first third-party measured humanoid results under a six-criteria benchmark. The Unitree G1 walked at 0.49 m/s in normal mode and 0.84 m/s in fast mode; a typical human walks about 1.4 m/s. A 3-kg payload did not slow the robot’s walking speed but added tenths of a second to acceleration. Fraunhofer plans to test additional platforms, making this the closest thing to an emerging independent test program.[1]

Everyday task success, measured on real hardware (Humanoid Everyday, Oct 2025) An academic benchmark of 260 everyday tasks run on real robot hardware found roughly 51% average task success across baseline systems, and 0% success on high-precision insertion tasks such as fitting a part into a tight slot. The pair of numbers quantifies the current state: robots succeed about half the time on structured everyday tasks and fail completely at the fine-motor edge cases. This is the hand problem, described below, expressed as data.[2]

The measurement layer is being built now (May–Jun 2026) NIST proposed the first standardized U.S. humanoid performance benchmark since the 2015 DARPA Robotics Challenge, covering baseline locomotion and manipulation; no cross-robot results have been published yet. RLWRLD and NVIDIA launched DexBench, a dexterity benchmark and data standard, on June 9, 2026. The AGIBOT World Challenge ran alongside ICRA 2026 in Vienna with 526 teams from 27 countries and a planned online simulation leaderboard, though it compares research teams’ control policies rather than commercial products. Each of these will generate comparable time series over the next one to two years; until then, this page reports measured results individually.[3]

▶ LLMs, World Models, and VLA Models (click to expand)

Large Language Models (LLMs) dominate the field of frontier model AI. The most advanced AI models (ChatGPT/Codex, Claude, Gemini, etc.) use the LLM framework and have almost infinite digital training materials, especially as social media and other digital platforms create more and more new training material every minute of every day. This training method is not ideal for physical AI training because it lacks visual and tactile reference points.

World models are neural networks trained on video and image data to simulate physics, gravity, object permanence, and cause-and-effect which allow robots to learn from watching rather than doing, potentially bypassing the data bottleneck. NVIDIA’s Cosmos platform, World Labs ($1B raised at $5.4B valuation, Feb 2026), and Yann LeCun’s AMI Labs ($1.03B raised at $3.5B valuation) are all building world models for robotics.[4]

Vision-language-action (VLA) models such as Physical Intelligence’s π0 (essentially “GPT for robots”) can be fine-tuned to new tasks with as little as 1–20 hours of demonstration data, rather than requiring thousands of hours of programming for each new capability.[5] The dropdown below profiles the leading VLA models in detail.

▶ Foundation Models for Physical AI: “GPT for Robots” (click to expand)

Just as GPT and Claude transformed language processing by training a single model on vast text data, a new class of vision-language-action (VLA) models aims to create a universal “brain” for robots. These models see the environment (vision), understand instructions (language), and output motor actions (action). If they succeed at scale, they decouple robot software capability from robot hardware development: a breakthrough in the foundation model could make every existing robot body dramatically more capable overnight, the way a software update transforms a smartphone without changing its hardware.

Model / Platform	Developer	Key Capability	Status (early 2026)
π0 (“pi-zero”)	Physical Intelligence (SF); $600M raised Nov 2025 at $5.6B valuation	Cross-embodiment generalist: pre-trained on 10,000+ hrs from 7 robot types and 68 tasks. Fine-tunes to new tasks with 1–20 hrs of data. Folds laundry, assembles boxes, routes cables. Open-sourced Feb 2025.	Version 0.6 (early 2026) with reinforcement learning doubled task throughput. Backed by Bezos, Sequoia, CapitalG. Closest to a “GPT moment” for robotics.
NVIDIA Isaac GR00T N1 (updated to N1.6)	NVIDIA; open-source	Dual-system architecture: “System 1” (fast reflexive action via diffusion transformer) + “System 2” (deliberate planning via vision-language model). Trained on human demos + massive synthetic data from NVIDIA Omniverse. Downloaded 1M+ times.	Available now. Early access partners: Agility, Boston Dynamics, Figure AI, 1X, Sanctuary AI. Foxconn deploying GR00T-powered humanoids at Houston factory Q1 2026. Newton physics engine co-developed with Google DeepMind and Disney Research.
RT-2 / RT-2-X	Google DeepMind	The original VLA: co-trained on robot data and web data, treating robot actions as text tokens. Emergent capabilities: spatial reasoning, improvised tool use (picks up a rock when asked for a hammer), multi-stage semantic reasoning. RT-2-X trained on Open X-Embodiment data showed 3× improvement in emergent skills.	Succeeded by Gemini Robotics (2025), built on Gemini 2.0 family with on-device versions. Google also maintains Open X-Embodiment, the field’s largest shared dataset (1M+ trajectories, 22 robot types, 33 institutions).
Helix	Figure AI; proprietary	“System 1 / System 2” VLA for humanoid whole-body control. Vertically integrated: Figure builds both brain and body (Figure 03). Designed for home + commercial environments.	Powers Figure 03; deployed at BMW. BotQ factory in Austin (12K initial capacity, scaling to 100K). Not open-source: competitive moat strategy.
OpenVLA / SmolVLA	Stanford (OpenVLA); Hugging Face (SmolVLA)	Open-source alternatives. OpenVLA: 7B-parameter model trained on ~970K episodes from Open X-Embodiment; often outperforms RT-2. SmolVLA: ~450M parameters, runs on consumer hardware (laptop GPU). Democratizes access.	Active research tools. Important for ecosystem development but not yet at the scale of π0 or GR00T for commercial deployment.

The foundation model landscape for physical AI in early 2026 mirrors the LLM landscape circa 2020–2021: multiple competing architectures, rapid iteration, no clear winner, and a critical dependence on training data scale. The key open question is whether these models can achieve the same kind of emergent generalization that LLMs demonstrated, where scaling data and compute produces capabilities that were not explicitly programmed. If they can, the humanoid timeline compresses dramatically. If they cannot, each new task and environment will continue to require expensive, manual data collection and fine-tuning.

Robotics & Physical AI: Binding Constraints[6]

Why is physical AI years behind software AI? The constraints are not primarily computational; the same GPU clusters that train language models can train robot policies. The constraints are physical and institutional, and they cluster around three core problems.

1. Dexterous Manipulation: “The Hand Problem” The human hand has 27 degrees of freedom, 34 muscles, and over 100 ligaments. The best humanoid hands in 2026 have 16–22 degrees of freedom (Figure 02: 16 DoF; Tesla Optimus Gen 3: 22 DoF with 50 actuators). Folding fabric, plugging cables, handling soft or irregular objects, and performing fine motor tasks in unpredictable environments remain beyond current systems. Tactile sensing (GelSight, MIT) and diffusion policy approaches (Columbia/MIT) are showing progress, but the gap between a robot that can grasp a rigid box and one that can fold a fitted sheet is enormous. This is the single hardest unsolved problem in physical AI and the primary reason humanoids are limited to structured tasks like tote handling.

2. Training Data for Physical Tasks Large language models were trained on the entire text of the internet, trillions of words, freely available. Physical AI requires data about real-world interactions: how objects move, how forces transfer, how materials deform. This data is expensive, slow, and sometimes dangerous to collect. Each new robot body requires new data. The largest public robotics dataset (Open X-Embodiment, Google DeepMind + 33 institutions) contains roughly 1 million robot demonstrations across 22 robot types, orders of magnitude smaller than LLM training corpora. Synthetic simulation data helps but the “sim-to-real gap” remains: behaviors that work in simulation often fail in the physical world. DoorDash, Uber, Sunday Robotics, and Instawork are now paying gig workers to generate this data, and Bessemer estimates the industry will spend $3B+ on robotic data within two years.

3. Battery Life and Endurance A human warehouse worker sustains an 8-hour shift (480 minutes). Agility’s Digit operates in 30-minute work intervals at Amazon warehouses. Figure 02 runs 2–3 hours. Sanctuary AI’s Phoenix demonstrated 43.5 cumulative hours at Hannover Messe 2025, but with charging breaks. The physics of bipedal balance consumes disproportionate energy relative to productive work. Current battery technology does not support full-shift autonomous operation for bipedal humanoids, and hot-swap battery systems are still early (UBTECH claims autonomous battery swap on Walker S2 but it is unverified at scale). Until humanoids can work a full shift, their unit economics cannot compete with human labor on a per-hour basis.

What Has Been Solved: Not all physical AI problems remain open.

Bipedal locomotion on structured surfaces (factory floors, flat warehouses) is largely solved; Boston Dynamics’ Atlas can run, jump, and perform dynamic maneuvers, and multiple commercial platforms navigate indoor environments reliably.
Pattern recognition and perception in controlled settings have improved dramatically through foundation models like RT-2 and π0, enabling robots to identify and interact with novel objects without explicit programming.
Communication between robots and human operators through natural language is functional, powered by the same LLM architectures that drive ChatGPT and Claude.

The unsolved problems are primarily about physical interaction with unpredictable environments, the domain where the real world is most different from the digital one (see LLMs, World Models, and VLA Models above).

Market Forecasts and Timeline Estimates

Side-by-side dot plots comparing AI automation vulnerability scores across occupation categories. Business and Finance jobs cluster between 0.4 and 0.7 with a mean of 0.49, while Production, Construction and Transportation jobs cluster near zero with a mean of 0.07

AI Automation Vulnerability by Occupation Category | Adapted from Srinivasan, Chen & Zakerinia, “Displacement or Complementarity? The Labor Market Impact of Generative AI” (HBS Working Paper, Dec 2024 / updated Aug 2025). Original data: 19,000+ tasks across 900+ U.S. occupations scored using OpenAI ChatGPT. Job postings 2019–Mar 2025.

Knowledge work appears to be more exposed to AI automation than physical work based on currently available technology. An HBS study scoring 19,000+ tasks across 900+ U.S. occupations found that business and finance roles average 0.49 on its automation index, while production, construction, and transportation roles average 0.07.[7] The reason is the set of physical constraints described above: dexterity, training data, battery life, safety standards, and maintenance infrastructure.

The chart above illustrates the high potential exposure business and finance jobs have to AI based automation compared to the much lower current exposure levels for more physically focused jobs in the production, construction, and transportation sectors.

What the HBS score does and does not measure. The study scores each occupation’s tasks on whether generative AI software can perform them. A model’s outputs are text, code, images, and analysis, so a task like preparing financial statements scores high, while a task like laying a weld bead scores near zero: no software output can actuate in the physical world. The score is therefore a faithful measure of software exposure, not a general measure of automation risk. Three occupations make the distinction concrete. Welders score 0.04 on the HBS index despite being the most thoroughly automated occupation of the twentieth century, with robotic arms welding in every major auto plant since the 1980s. Industrial truck operators score 0.00 while automated guided vehicles displace forklift drivers in warehouses today. Stockers and order fillers score 0.33, well above welders, because the index detects the clerical fraction of their work, the inventory records and scheduling, not the physical picking.

This is Moravec’s Paradox in a 2025 dataset: tasks humans find hard, such as tax analysis, are easy for AI, while tasks a toddler finds easy, such as picking up an oddly shaped object, remain hard for machines. The software wave and the physical wave are different technologies on different timelines. This page covers the second one, and the constraints and measurements above describe where it actually stands.

Select Estimates and Projections – Robotics and Physical AI Automation

Source	What They Estimate	Market / Shipment Figures	Labor Impact Estimate	Horizon	Key Caveat
HBS (Srinivasan et al., Dec 2024 / Aug 2025)	AI automation vulnerability by occupation (19,000+ tasks, 900+ occupations)	N/A	Post-ChatGPT: automation-prone job postings −13%; augmentation-prone postings +20%. Business & finance automation score: 0.49 mean. Production/construction/transport: 0.07 mean.	2019–Mar 2025	Empirical observation, not forecast. Measures actual job posting changes. Focuses on generative AI (software), not physical robotics. Physical work scores near zero because the technology to automate it at scale doesn’t exist yet.
iCapital / Multi-Bank Consensus (Aug 2025)	Humanoid adoption & TAM (average of BofA, Citi, Morgan Stanley, UBS base cases)	TAM ~$4.5T by 2050; ~1M units by 2030; ~25M by 2035; ~800M by 2050. BoM ~$40K by 2026, ~$10K by 2040.	5 in 10 U.S. manufacturing positions expected vacant through 2033; China faces 22% labor force decline by 2050. Humanoids could boost manufacturing efficiency 20–30% in 5 years.	2025–2050	Consensus average across four major banks. Three adoption phases: industrial (2025–2030), services & healthcare (2031–2035), household/societal (2036+). VC: $3.1B in H1 2025 > $2.9B entire 2010–2024 period. ~10% U.S. household adoption by 2050.
Goldman Sachs (Jan 2024 + Mar 2026)	Humanoid robot market; AI labor impact	$38B market by 2035; 250K+ units by 2030; 1.4M units by 2035; 70% CAGR	6–7% of U.S. workers displaced over ~10 years; AI can automate 25% of U.S. work hours; 0.6pp unemployment increase if adoption spread over a decade[8]	2024–2035	Market estimate revised 6× upward from prior $6B estimate after costs declined 40%. Mar 2026 update: displaced knowledge workers may be poorly suited for the labor most needed (HVAC, electricians, construction). Data center build-out has already created 216,000 construction jobs since 2022.
Morgan Stanley (Apr 2025)	Total humanoid ecosystem (hardware + supply chain + services)	$5T by 2050; 1B+ units; 13M units by 2035	U.S.: 8M humanoid workers by 2040 ($357B wage impact); 63M by 2050 ($3T payroll, 75% of occupations, 40% of employees affected)	2025–2050	25-year forecast. Adoption “relatively slow” until 2035, then accelerating. $5T includes entire ecosystem. Consumer home applications a decade away. 10% of U.S. households may own a humanoid by 2050.
IFR (Sep 2025)	Industrial robots (all types, not just humanoid)	542K annual installs (2024); 4.66M operational stock; 575K projected for 2025; 700K+ by 2028	N/A: IFR reports installations, not labor impact	Actual through 2024	Measured, not forecasted. Humanoids are ~3% of annual volume. This is the installed base doing real work in factories today.
Counterpoint Research (Jan 2026)	Humanoid robot shipments	~16K units in 2025; China = 80%+; 100K+ cumulative by 2027; 69.7% CAGR to 2030	N/A	2025–2030	Most 2025 units are developer kits, research platforms, and entertainment deployments, not robots doing productive autonomous work. “Units shipped” ≠ “units deployed productively.”
ABI Research (Jul 2025)	Humanoid market size & units	$6.5B by 2030; 138% CAGR; 115K units by 2027	N/A	2024–2030	More conservative near-term than Goldman. Market “heats up” in 2027, not 2025–2026. Inflection depends on regulatory, safety, and ROI issues being resolved.
Deloitte (Nov 2025)	Industrial robot installed base (all types)	5M+ cumulative by 2025; 5.5M by 2026	N/A	2025–2026	Conservative. Growth stays “relatively modest” without solving data quality, integration, and cybersecurity bottlenecks.

What to Watch in 2026

Subject	Key Source(s)	Expected Next Release
Humanoid robot market forecasts	Goldman Sachs Research	Periodic updates; last major labor report Mar 2026. Watch for revised humanoid market estimates based on 2025 actual data
Physical AI foundation models	Physical Intelligence (π) / NVIDIA Cosmos	Continuous; watch for cross-embodiment generalization milestones and commercial licensing
Independent humanoid benchmarks (new field)	NIST baseline benchmark; Fraunhofer IPA test program; DexBench (RLWRLD + NVIDIA); AGIBOT World Challenge sim leaderboard	All launched May–Jun 2026. Watch for first NIST cross-robot results and Fraunhofer tests of additional platforms; these become the field’s first comparable time series
Robotic training data investment	DoorDash Tasks, Uber AI Solutions, Scale AI	Ongoing; watch for scale of gig-worker data programs and whether data costs decline as world models mature

[1] Fraunhofer IPA humanoid robot benchmark, first published results (May 2026): six application-relevant criteria; Unitree G1 measured walking speeds 0.49 m/s (normal) and 0.84 m/s (fast); 3-kg payload did not reduce walking speed but slowed acceleration by tenths of a second. Human reference walking speed ~1.4 m/s. ipa.fraunhofer.de ↩

[2] Humanoid Everyday benchmark (arXiv, Oct 2025): 260 everyday tasks evaluated on real robot hardware; approximately 51% average task success across baseline systems; 0% success on high-precision insertion tasks. ↩

[3] NIST proposed baseline humanoid performance benchmark (May 2026), the first U.S. humanoid standardization effort since the 2015 DARPA Robotics Challenge. RLWRLD + NVIDIA DexBench launch (Jun 9, 2026). AGIBOT World Challenge 2026 at ICRA Vienna (Jun 2026): 526 teams from 27 countries; online simulation evaluation plus real-robot finals; public simulation leaderboard planned. ↩

[4] World Labs raised $1B at $5.4B valuation (Feb 2026); AMI Labs (Yann LeCun) raised $1.03B at $3.5B valuation (Mar 2026). Sources: Not Boring newsletter (Mar 2026); company announcements. NVIDIA Cosmos platform released progressively from CES 2025 through Feb 2026. ↩

[5] Physical Intelligence, “Our First Generalist Policy” (Oct 2024); open-sourced Feb 2025. $600M raised Nov 2025 at $5.6B valuation. The Robot Report (Nov 26, 2025). pi.website ↩

[6] Two additional constraints, regulatory/liability vacuum and maintenance/field support infrastructure, also impede deployment. No ISO standard exists for dynamically balancing legged robots working near humans, and the liability question (manufacturer vs. deployer vs. AI provider) is unresolved in all jurisdictions. Maintenance infrastructure (trained technicians, spare parts, 24/7 support) does not exist at scale for humanoids. These constraints are institutional rather than technical, but they may prove equally binding in practice. ↩

[7] Srinivasan, Suraj, Wilbur Xinyuan Chen, and Saleh Zakerinia. “Displacement or Complementarity? The Labor Market Impact of Generative AI.” Harvard Business School Working Paper (Dec 2024, updated Aug 2025). Study scored 19,000+ tasks across 900+ U.S. occupations using OpenAI ChatGPT to categorize automation vs. augmentation potential. Job postings 2019–Mar 2025. After ChatGPT’s launch, postings for automation-prone roles decreased 13%; postings for augmentation-prone roles grew 20%. Skills required for automation-prone roles shrank 7%. hbs.edu ↩

[8] Goldman Sachs Research, “How Will AI Affect the US Labor Market?” (Mar 18, 2026). Joseph Briggs, co-lead of the global economics team. Base case: 6–7% of workers displaced over ~10 years; 0.6pp unemployment increase; construction jobs exposed to data center build-out increased 216,000 since 2022; ~500,000 net new jobs needed for power demand by 2030. goldmansachs.com ↩

Sources

[1] Fraunhofer IPA, humanoid robot benchmark first results (May 2026). ipa.fraunhofer.de

[2] Humanoid Everyday benchmark (arXiv, Oct 2025).

[3] NIST baseline humanoid benchmark (May 2026); RLWRLD + NVIDIA DexBench (Jun 9, 2026); AGIBOT World Challenge 2026, ICRA Vienna.

[4] Physical Intelligence, π0 (Oct 2024; open-sourced Feb 2025; $600M Nov 2025). pi.website

[5] NVIDIA, Isaac GR00T N1 (Mar 2025; N1.6 2026) and Cosmos; World Labs and AMI Labs (company announcements, Feb–Mar 2026).

[6] Open X-Embodiment Collaboration / Google DeepMind, RT-2 / RT-X (2023, updated 2025).

[7] Srinivasan, Chen & Zakerinia, “Displacement or Complementarity?” HBS Working Paper 25-039 (Dec 2024, updated Aug 2025). hbs.edu

[8] Goldman Sachs Research, “How Will AI Affect the US Labor Market?” (Mar 18, 2026); “Humanoid Robot: The AI Accelerant” (Jan 2024). goldmansachs.com

Companion workbook. technology-ai-blue-collar.xlsx: shared robotics and physical AI workbook covering the technical constraints scorecard, occupation automation risk matrix, market-forecast and company-landscape data, and robot installation and density series