howclose.to
Intelligence & Machines · Updated July 2026Momentum · accelerating

How close are we to general-purpose robotics?When will robots handle everyday work?

or, simply: When will robots handle everyday work?or, precisely: How close are we to general-purpose robotics?

Legged locomotion is essentially solved, but dexterous manipulation and open-world autonomy are not, so today's factory and home robots run narrow, supervised, and heavily teleoperated.Robots can walk, run and keep their balance beautifully now. The hard part left is hands and judgment: doing varied tasks all day, safely, without a human quietly steering.

We are here

Figure 03 deploys for logistics at BMW Spartanburg - After the Figure 02 pilot, BMW put Figure 03 to work on complex logistics sequencing at Spartanburg, an expanded but still narrow, supervised deployment. Next up - First Neo units ship to homes (expected 2026).

01 · Where we stand

State of playWhere general robotics stands right now

The current stage, the honest metric, and the single threshold that gates the next stage. Each threshold is a falsifiable claim with a named next test.How far up the ladder we've climbed, the honest verdict, and the one thing blocking the next step.

The five stagesMaturity ladder
Deployed
Out in the real worldDeployed at scale
Scaling
Making it cheap enough at scaleScaling toward competitive cost
Engineering
Building one that pays for itselfEngineering a system that pays back
Lab demo← HERE
Shown to work in a labDemonstrated in the laboratory
Theoretical
The idea is worked out on paperTheoretical basis established
The honest verdictVerified state

Robots can walk, run and keep their balance beautifully now. The hard part left is hands and judgment: doing varied tasks all day, safely, without a human quietly steering.Legged locomotion is essentially solved, but dexterous manipulation and open-world autonomy are not, so today's factory and home robots run narrow, supervised, and heavily teleoperated.

No single honest number for this one - read where it stands by the milestones and the next test.No single honest scalar - progress is read by milestones and the named next test, not a headline figure.

Blocking the next stepBlocking threshold

Handle unfamiliar objects and toolsDexterous manipulation Next test: Independent long-horizon manipulation trials on tools and deformable objects, e.g. pi-0.6 field runs

The verdict, in five rungsHow far up the ladder↓ next — the thresholds & the gap
01 · The evidence

The thresholds that gate the next stageWhat has to happen next

Each threshold is a falsifiable claim with a named next test; the gap chart shows how far today's metric sits from the goal.Each row is one thing that has to be proven — and how far today's number is from the target.

Dynamic locomotion & balanceWalk, run and keep balance like an animal✓ Achieved · 2024
100%
Proven byRL whole-body control; Atlas, Unitree and others walk, run and recover from shoves
Dexterous manipulationHandle unfamiliar objects and toolsEarly
30%
Next testIndependent long-horizon manipulation trials on tools and deformable objects, e.g. pi-0.6 field runs
Open-world autonomyWork for hours without a human steeringEarly
15%
Next testPublished intervention-rate and uptime data from paid deployments
Labor economicsCost less than the work is worthEarly
12%
Next testMulti-site customers renew deployments without vendor operators on site
THRESHOLDS - Thresholds for General Robotics.
02 · How we got here

The record behind the verdict

Major events set large; context events set small but never hidden. Everything below the TODAY rule is a schedule, not a result.

1961-19991 event1 shown

Fixed industrial arms

Fixed industrial arms begins with unimate, the first industrial robot, joins a gm line. The result established the next question for the field.

1961
Unimate, the first industrial robot, joins a GM lineDeployment
Unimation's hydraulic Unimate arm began unloading a die-casting machine at a General Motors plant, launching factory robotics.
2000-20202 events1 shown

Dynamic legged locomotion

Dynamic legged locomotion moved the field from honda unveils asimo to darpa robotics challenge finals (and the falls). The results narrowed the next question without closing it.

2000
Honda unveils ASIMOExperiment
Honda revealed ASIMO, a 120cm bipedal humanoid that could walk, climb stairs and recognize faces, becoming the era's icon of legged robotics.
2015
DARPA Robotics Challenge Finals (and the falls)
At the Pomona finals most humanoids struggled with doors, valves and rubble, and a widely shared reel of robots toppling over exposed how brittle whole-body autonomy still was.
2021-20244 events2 shown

The vision-language-action turn

The vision-language-action turn moved the field from boston dynamics atlas runs parkour to open x-embodiment pools cross-robot data. The results narrowed the next question without closing it.

2021
Boston Dynamics Atlas runs parkourExperiment
Two Atlas robots vaulted, leapt gaps and did backflips through a parkour course, showcasing dynamic whole-body control (though heavily choreographed and hydraulic).
2022
Tesla shows first Optimus prototype
At AI Day a rough Optimus prototype walked onstage and waved; Tesla framed it as an early development robot, kicking off the commercial humanoid race.
2023
Google DeepMind's RT-2 vision-language-action modelTheory
RT-2 fused web-scale vision-language pretraining with robot data so a single model could turn natural-language commands into actions and generalize to unseen objects, opening the VLA era.
2023
Open X-Embodiment pools cross-robot data
A shared corpus combined more than one million real-robot trajectories from 60 datasets across 22 embodiments, enabling the first major cross-robot training effort.
2025-202714 events3 shown

General-purpose humanoids

General-purpose humanoids moved the field from pi-0 code and weights released to first neo units ship to homes. The results narrowed the next question without closing it.

2025
pi-0 code and weights released
Physical Intelligence published pi-0's code, weights and fine-tuned checkpoints, making its cross-embodiment VLA reproducible outside the originating lab, with explicit transfer limitations.
2025
Helix controls a humanoid upper body
One VLA controlled a humanoid's torso, wrists, head and fingers and coordinated two robots without task-specific fine-tuning; this was a vendor lab demonstration.
2025
Gemini Robotics adds direct actions
Google DeepMind added robot actions to Gemini 2.0 and demonstrated language-steerable manipulation across ALOHA, Franka and an adapted Apollo humanoid in lab evaluations.
2025
GR00T N1 released as an open model
NVIDIA released an open-weight humanoid foundation model trained on human videos plus real, simulated and synthetic robot trajectories; results remained developer-evaluated.
2025
Physical Intelligence pi-0.5 generalizes to new homesTheory
pi-0.5 extended the VLA approach with open-world generalization, letting a robot clean kitchens and bedrooms in homes it had never seen during training.
2025
Gemini Robotics runs on-device
An on-device VLA ran locally and opened to trusted testers, with adaptation demonstrated from 50 to 100 examples; this was not a paid-deployment reliability result.
2025
1X opens orders for the Neo home robot (~$20k)Deployment
1X began selling the Neo humanoid for $20,000; crucially, tasks Neo can't do autonomously are performed by a remote human teleoperator who can see into the home.
2025
Digit passes 100,000 totes moved
Agility reported more than 100,000 totes moved at GXO's Flowery Branch facility, cumulative throughput for a constrained workflow without published intervention rate or robot-hours.
2026
pi-0.6 reaches customer-site tasks
A packaging robot completed a partner-reported shift at 96.4% autonomy while laundry-folding systems worked in commercial laundromats; both remained narrow, human-supported workflows.
2026
VLA memory reaches 15-minute tasks
Multi-scale Embodied Memory demonstrated kitchen and meal-preparation tasks requiring up to 15 minutes of context, still far short of unsupervised work-shift autonomy.
2026
Gemini Robotics-ER 1.6 opens by API
The embodied-reasoning component became API-accessible for spatial reasoning, task planning, success detection and instrument reading; it was not an end-to-end robot deployment.
2026
pi-0.7 recombines manipulation skills
One generalist policy recombined learned skills across robots, scenes and tasks without specialist fine-tuning; the evidence was an internal evaluation, not independent open-world validation.
2026
Figure 03 deploys for logistics at BMW SpartanburgDeploymentWe are here
After the Figure 02 pilot, BMW put Figure 03 to work on complex logistics sequencing at Spartanburg, an expanded but still narrow, supervised deployment.
2026
First Neo units ship to homesDeploymentTarget
1X plans to begin delivering Neo to early-access buyers during 2026, one of the first humanoids in ordinary households, initially leaning heavily on teleoperation.
2024-20276 events1 shown

Events outside the declared eras

Events outside the declared eras moved the field from digit enters a multi-year warehouse deployment to tesla's stated goal: mass humanoid manufacturing. The results narrowed the next question without closing it.

2024
Digit enters a multi-year warehouse deployment
GXO signed a Robots-as-a-Service agreement for Digit robots moving totes at a live SPANX facility; the achieved deployment covered one constrained workflow.
2024
Unitree G1 drops to $16,000
Unitree launched the G1 humanoid at a base price of $16,000, an order-of-magnitude cut that put walking humanoids within reach of labs and small firms.
2024
Figure 02 pilots at BMW SpartanburgDeployment
Figure 02 ran a real production pilot at BMW's South Carolina plant, inserting sheet-metal parts into fixtures; it was a limited, closely supervised trial rather than full autonomy.
2024
Physical Intelligence releases pi-0
Physical Intelligence's pi-0 generalist policy ran one model across multiple robots to fold laundry, bus tables and bag groceries, demonstrating cross-embodiment dexterous skills.
2024
4 million industrial robots now in factories
The IFR logged a record 4.28 million operational industrial robots worldwide, the fixed-arm base on which the humanoid wave is building.
2027
Tesla's stated goal: mass humanoid manufacturingFundingTarget
Tesla says a next-generation line is designed for up to 10 million Optimus units per year at a ~$20k target cost; this is an aspirational goal, not a demonstrated capability (Tesla has missed every prior Optimus timeline).
— end of record · 27 shown, 0 hidden —
27 events · below the TODAY rule = scheduled, not done
03 · The data behind the verdict

Why the meters read the way they do

The learning curves and comparisons that justify each threshold's percentage. Every series is measured, with the source event linked in the timeline above.

Movement is banked. Useful work is not.1 of 4 capability gates are demonstrated; Dexterous manipulation is the first open test.One marker = one capability gate · solid = demonstrated · hollow dashed = unproven · locomotion → useful workOne marker = one capability gate · verified gates are solid; open gates are hollow and dashed · ordered from locomotion to fully loaded labor economics
  1. BANKED · 2024

    Walk, run and keep balance like an animalDynamic locomotion & balance

    Getting around on two legs used to be the famous hard problem. Reinforcement learning has mostly cracked it.Robust bipedal walking, running and whole-body recovery over uneven terrain

    OBSERVED EVIDENCERL whole-body control; Atlas, Unitree and others walk, run and recover from shoves

  2. CURRENT FRONTIER

    Handle unfamiliar objects and toolsDexterous manipulation

    Robots grab an apple almost every time, but a spoon, scissors or a fitted part still trips them up.Transfer fine-motor skills across unfamiliar objects, tools and cluttered scenes

    DECISIVE NEXT TESTIndependent long-horizon manipulation trials on tools and deformable objects, e.g. pi-0.6 field runs

  3. DOWNSTREAM GATE

    Work for hours without a human steeringOpen-world autonomy

    Home and factory robots still lean on remote operators and reset every 30-90 minutes. Boring, reliable shifts are the goal.Complete multi-hour tasks safely with rare human intervention or teleoperation

    DECISIVE NEXT TESTPublished intervention-rate and uptime data from paid deployments

  4. DOWNSTREAM GATE

    Cost less than the work is worthLabor economics

    A $20k price tag ignores the teleoperators, supervision and repairs that make a robot-hour actually work.Fully-loaded cost per useful hour competitive with human labor

    DECISIVE NEXT TESTMulti-site customers renew deployments without vendor operators on site

BOSTON DYNAMICS · UNITREE · PHYSICAL INTELLIGENCE · AS OF JUL 2026
Technical notes

Read the evidence more closely

Definitions, system boundaries and experimental caveats behind the headline record.

01Cross-robot gains are benchmark-relative

Open X-Embodiment combined 60 datasets from 34 labs and 22 embodiments. RT-1-X improved performance by 50% in small-data evaluations, while RT-2-X scored 3× RT-2 on the project's emergent-skill evaluation. These are relative benchmark results, not absolute reliability rates.

02Open π₀ still needs local robot data

Open π₀'s authors report that adapting it to individual tasks generally required 1-20 hours of robot data, but explicitly warn that adaptation to outside platforms may fail.

03Helix splits reasoning from control

Helix separates slower semantic reasoning from fast motor control: its onboard VLM runs at 7-9 Hz, while the action system handles high-rate continuous upper-body control. “Pick up virtually anything” remains Figure's wording, not an independently bounded success rate.

04Gemini's gains lack deployment measures

Gemini Robotics reportedly more than doubled average performance over previous VLAs on DeepMind's generalization benchmark; Gemini Robotics-ER achieved 2-3× Gemini 2.0's success rate in an end-to-end internal setup. Neither result supplies paid-deployment uptime or intervention data.

05On-device transfer followed adaptation

Gemini Robotics On-Device's nine-task internal evaluation reported roughly 0.52-0.74 success rates, versus approximately 0.11-0.36 for the previous on-device baseline. It was trained principally for ALOHA and then adapted to Franka and Apollo, so this is transfer after adaptation-not zero-shot embodiment interchangeability.

06Synthetic-data gains are model-level

NVIDIA generated 780,000 synthetic trajectories, described as equivalent to 6,500 demonstration-hours, in 11 hours; mixing them with real data improved GR00T N1 by 40% against NVIDIA's real-data-only baseline. This is an internal model result, not a 40% improvement in deployed labor productivity.

07Digit's deployment metrics remain sparse

Digit's 100,000-tote result demonstrates repeated production cycles, but the published workflow remains tote pickup, carrying and placement integrated with AMRs and conveyors. Agility does not publish intervention rate, robot-hours, fleet size, throughput per robot or cost per tote.

08Autonomy still includes human recovery

In the π₀.6 partner results, incorporating deployment-specific laundry data reduced missed-grasp sequences by 42% and interventions by 50%. Human-in-the-loop operators still handle failures, making intervention frequency-not binary “autonomous” labeling-the important variable.

09Fifteen-minute memory is not a work shift

Multi-scale Embodied Memory retains recent observations through a video encoder but compresses long-term state into natural-language memories selected by the model. Its demonstrated horizon is up to 15 minutes, still well short of an unsupervised work shift.

10Instrument reading is a component test

Gemini Robotics-ER 1.6's internal instrument-reading score rose from 23% for ER 1.5 to 86%, or 93% with agentic vision. This evaluates perception and reasoning over instrument images; it does not demonstrate autonomous physical inspection end to end.

11π₀.7 separates heterogeneous behaviours

π₀.7's key mechanism is multimodal conditioning-language, strategy metadata, desired episode duration and visual subgoals-which lets heterogeneous human, robot and autonomous-experience data coexist without conflating different behaviours. Its claimed compositional generalization remains developer-evaluated.

04 · What it unlocks

If the remaining tests pass

Downstream capabilities, drawn dashed because they depend on results not yet in.

General RoboticsSafer dangerous workmachines take more jobs in mines, fires and disaster zonesAbundant physical laborbuilding, moving and maintaining things gets much cheaperIndependent living supportpractical daily help reaches more older and disabled people
05 · The players & the money

Who is building it-and what the money saysCapital, institutions and the global race

The teams doing the work, where they are based, and whether the money points to real delivery or only a plan.Company finance, public programmes, institutional leadership and market evidence-kept separate from valuations, forecasts and announced capacity.

The short versionCapital and institutional readout

General-purpose robotics is drawing large rounds, but commercial evidence is still dominated by narrow material-handling and manipulation workflows. US labs lead frontier robot policies, China leads manufacturing scale, and incumbent industrial automation remains far larger than the humanoid sector.

Players

Who is building itCompanies, laboratories and programmes

Figure AI

USA

Humanoid hardware, Helix vision-language-action models and factory/logistics deployments.

Apptronik

USA

Apollo humanoid for manufacturing and logistics, with Google DeepMind model collaboration.

Agility Robotics

USA

Digit humanoid and Arc fleet platform for repetitive warehouse movement.

Companies, laboratories and programmes working on this problem
PlayerCountryWhat they are doingFundingNamed investorsSource
Figure AIcompanyUSA

Humanoid hardware, Helix vision-language-action models and factory/logistics deployments.

The valuation is not funding, and announced factory capacity is not achieved robot output.

Series C · >$1B · Sep 2025 · $39B post-money valuationParkway Venture Capital · Brookfield · NVIDIA · Macquarie Capital · Intel CapitalSource · figure.ai
ApptronikcompanyUSA

Apollo humanoid for manufacturing and logistics, with Google DeepMind model collaboration.

Series A · $403M · Mar 2025B Capital · Capital Factory · Google · Mercedes-Benz · Japan Post CapitalSource · apptronik.com
Agility RoboticscompanyUSA

Digit humanoid and Arc fleet platform for repetitive warehouse movement.

A proposed SPAC valuation is not treated as a completed funding round.

Series B · $150M · Apr 2022DCVC · Playground Global · Amazon Industrial Innovation Fund · Sony Innovation FundSource · agilityrobotics.com
Physical IntelligencecompanyUSA

Cross-embodiment pi-series foundation policies and deployment-specific adaptation.

Funding is omitted because no primary company announcement verifying reported round totals was found.

Not disclosedNot disclosedSource · pi.website
Google DeepMind RoboticslabUK / USA

Gemini Robotics, embodied reasoning and Open X-Embodiment research.

Funded inside Alphabet; no standalone lab financing round.

Not disclosedNot disclosedSource · deepmind.google
NVIDIA RoboticslabUSA

GR00T open models, simulation, synthetic data and robot-compute platforms.

NVIDIA corporate financing is not represented as robotics-lab funding.

Not disclosedNot disclosedSource · research.nvidia.com
06 · Sources

Where every number comes from

4 sources — every figure on this page traces to one.