The bear case is the argument that robotics is repeating earlier AI winters: capital arrives on demonstrations, deployment evidence arrives late, and funding runs out before the reliability gap closes. Robotics startups raised $18.8B by late June 2026 by Crunchbase's count, past 2025's record $15BSource 9.

Honda's white ASIMO humanoid robot standing on a polished floor beneath the large suspended globe display at Miraikan, Tokyo.
Honda's ASIMO performing at Miraikan, Tokyo, in 2016. The bear case holds that capital arrives on demonstrations like this one and deployment evidence arrives late.
Photo: Sylvain Kalache / Wikimedia Commons. Licence: CC BY 2.0.

Eight milestones would settle the argument. Under "Show", "Done" empties the board and "Partly" leaves five tiles. In one, RL-100 and π*0.6 report some of the highest published per-task reliabilitiesSource 12 Source 13.

Bear-case milestones
Eight results would prove skeptics wrong; none has happened yet
Show
Adjust more
0 of 8 done

5 partly there, 3 not yet, 0 done

Showing all 8

Watch these eight: as of September 2026 five are partly there, and each one met would weaken the case that robotics is overhyped.
How this was made
Each status is Robot Wiki’s editorial call on published evidence, reviewed in September 2026: “partly there” means a published result moves toward the milestone without meeting its test; “not yet” means no inspected record does. The table gives each milestone’s original name, why it matters, the evidence behind the call with its sources, and how we’d know it had been met.
Eight milestones that would settle the bear case, with the status call, the evidence behind it and the observation that would flip it to met.
MilestoneWhy it mattersStatusEvidenceHow we’d know
One policy, >90% in unseen homesOne robot brain succeeds in 9 of 10 new homesThe generalization test: a single policy succeeding across many homes it has never seen, with no per-site data, is the clearest signal that robot intelligence is actually general.Partly thereπ0.5 cleaned kitchens and bedrooms in homes that were entirely absent from its training data, and Figure reports 56% zero-shot success in 30 unseen homes against 9% without Index pretraining, a company figure with no published trial counts; neither result reaches 90% across more than ten unseen homes on standardized tasks. Source 11Source 29A published, reproducible evaluation of one policy across more than ten unseen homes with standardized task definitions and success above 90%.
Verified 10,000-unit deployment10,000 humanoids doing documented workCommercial viability at scale: five figures of humanoids doing documented productive work would end the pilot-program era.Not yetThe inspected company records do not document 10,000 humanoids doing productive work: Agility reports 65,000 hours of Digit production experience without a unit count, while Figure reports 1,250+ hours and 90,000+ parts at BMW. This is an evidence-status call for these records, not a census of every manufacturer. Source 4Source 5Company filings or independent reporting confirming 10,000 or more units in productive work, with task performance documented.
Open benchmark, cross-lab agreementLabs agree on one shared public testThe evaluation crisis: until labs publish comparable numbers on shared tasks, every demo is its own benchmark and progress claims cannot be arbitrated.Partly thereRoboArena crowd-sources double-blind pairwise evaluations across a distributed network of evaluators instead of standardizing fixed tasks, RoboChallenge runs a reproducible online evaluation system, and ManipulationNet delivers reproducible task setups through standardized hardware kits; none has yet produced comparable absolute success numbers across independent labs. Source 14Source 15Source 16Multiple independent labs publishing comparable results on the same benchmark, with agreement on what the numbers mean.
RL >99% across a broad task setPractice gets robots to 99% on many tasksThe reliability-gap test: reinforcement learning closes the gap to 99%+ one task at a time today; whether it does so across many tasks at once decides if the gap is engineering or science.Partly thereRL-100 reports 100% success across 1,000 evaluation episodes on eight tasks, and π*0.6 passed 90% on three production-style tasks; both produce per-task specialists, and neither claims transfer to unseen tasks. Generalist reports 99% average success for GEN-1, a company figure from a recipe that is not RL-specific, and says not all tasks reach that rate. Source 12Source 13Source 30Published results showing better than 99% success across more than twenty diverse tasks from a single training recipe.
Tactile model inside a VLA pipelineTouch sensing built into the robot’s brainThe direct test of Brooks's argument that vision-only training cannot produce dexterity: if touch is the missing channel, integrating it should move contact-rich performance.Partly thereSparsh-X (roughly a million contact-rich interactions) and TouchWorld are tactile foundation models, and Figure's Helix 02 takes fingertip tactile input. T-Rex, a VLA that reads high-frequency touch, beats its strongest baseline by over 30% on 12 contact-rich tasks, and removing touch drops its six-task average from 65% to 42%; TouchWorld reports that removing touch causes its largest drop. Both are single research benchmarks from mid-2026. Source 17Source 18Source 31Source 32A VLA trained with tactile input outperforming its vision-only counterpart on contact-rich tasks, published with ablations.
Contact-rich sim-to-real, zero real dataHand skills learned only in simulation work for realIf simulation fidelity reaches contact-rich manipulation, the data bottleneck breaks: training data becomes cheap, unlimited, and parallel.Not yetSim-to-real transfer is routine for locomotion but hardest where manipulation makes contact: reality-gap surveys name inaccurate contact modeling, where contact states alternate between sticking, slipping, and separation, as a key contributor to the dynamics gap. Source 19A policy trained entirely in simulation achieving above 90% on a real contact-rich task, replicated outside the originating lab.
Humanoid cost-per-task parityA humanoid matches a specialist machine’s cost per jobThe form-factor question in one number: if a humanoid's total cost per task matches a purpose-built system in the same application, the general-purpose thesis survives contact with accounting.Not yetThe inspected palletising-cell cost guide supplies budget and payback context for a task-specific application, not a head-to-head humanoid comparison. These materials cannot establish humanoid cost-per-task parity; absence from them is not proof that no comparison exists elsewhere. Source 35A published total-cost-of-ownership analysis showing humanoid cost per task at or below a purpose-built system for a specific application.
100x data, proportional gain100 times more data brings matching gainsThe scaling-law question: the bull case rests on robot performance scaling with data the way language did; a plateau would vindicate the structure-and-priors camp.Partly thereEgoScale fit a log-linear scaling law (R² = 0.9983) between egocentric pretraining hours and validation loss, and reports downstream task completion rising from 0.30 at 1k hours to 0.71 at 20k hours; Lin et al. find a roughly power-law relationship for single-task generalization, but no success-rate scaling law across unseen environments has been published. Source 20Source 33A systematic evaluation showing a 10 to 100x data increase producing proportional real-world success-rate improvement.
>$23B
raised by early June 2026
PitchBook tally, by early June
−52.55%
Unitree Q1 profit
adjusted net, year over year, Q1 2026
0 of 8
milestones met
the watchlist above, Sep 2026

Brooks calls the dexterity forecasts fantasy, and Morgan Stanley asks for ROI

Evidence: Rodney Brooks, co-founder of iRobot and Rethink Robotics, calls the belief that current methods will yield practical humanoid dexterity within decades "pure fantasy thinking", since vision-only training gives no force channel (dexterity)Source 1. His January 2026 scorecard predicts deployable dexterity "will remain pathetic compared to human hands" beyond 2036, and walking humanoids too unsafe near people without new mechanical systemsSource 2. Rethink raised $150 million and shut down in 2018 when sales fell shortSource 21.

A round grey and black iRobot Roomba 870 robot vacuum cleaner on a light wooden floor.
An iRobot Roomba 870 in 2014. Rodney Brooks co-founded iRobot, which makes the Roomba, so his forecasts come from someone who has shipped robots.
Photo: Kārlis Dambrāns / Wikimedia Commons. Licence: CC BY 2.0.

Evidence: Morgan Stanley's July 2026 note says humanoids have a "PR problem": they are framed as substitutes for workers instead of tools for hazardous or labor-constrained tasks. It adds that "the industry's social license to deploy may matter just as much as technical performance". Investors "are increasingly looking for tangible evidence of real-world return on investment". The note says the US ban on Chinese humanoid imports risks raising R&D costs, and keeps a 50,000-unit China shipment estimate for 2026Source 3.

Unitree's Q1 2026 adjusted profit fell 52.55% year over year

Evidence: Unitree's Q1 2026 adjusted net profit fell 52.55% year over year, according to the STAR Market prospectus it filed in May 2026Source 34.

Evidence: Bessemer expects industry data costs to exceed $3B over the next two years and calls capital the moatSource 6 (industrial deployment).

Evidence: at Computex 2026, a Qualcomm-powered humanoid collapsed face-first during the live keynote, in a failure attributed to a communication glitchSource 7.

Vendor hours and shipment counts do not document productive work

The bear case is that closing the reliability gap between demo and product outlasts the capital. Lisa Yan, quoted in Bessemer's own analysis, describes the climb from 99% to 99.9% as steepSource 6.

Agility reports 65,000 hours of Digit production experience on an undated homepageSource 4, and Figure reports 1,250+ hours at BMW Spartanburg by November 2025Source 5; neither counts autonomous units or supervised time (evaluation crisis). Omdia counted about 13,000 humanoids shipped in 2025Source 23 Source 22 and SAG 19.1K in the first half of 2026Source 24. The IFR counted about 7,000 full-size humanoids, those above 140 cm, sold in 2025Source 25. Tesla's Q2 2026 update says it is still installing its first Optimus lines, whose initial builds go to an "Optimus Academy for training data collection"Source 26.

Robotics funding passed 23 billion dollars by June 2026, and China repriced first

Evidence: beyond Crunchbase's narrower venture countSource 9, a newsletter writeup put the total above $23B by early June 2026 by PitchBook's tallySource 8. Bessemer Venture Partners argues there is "no robotics bubble" and that the sector is "structurally underinvested", but concedes that "some valuations are stretched"Source 6. The skeptic side rests on Brooks's timeline predictions and Morgan Stanley's demand for ROI evidence: Brooks calls some humanoid hype "downright delusional"Source 2, while Morgan Stanley makes no bubble callSource 3 (competing theses).

Robotics funding and humanoid counts
Startups raised a record while humanoids shipped in the thousands
Raised by robotics startups2020202120222023202420252026$18.8B to 22 June$0$10$20billions of US dollars a yearHumanoids countedFull-size humanoids sold in 2025 (IFR)about 7,000Humanoids shipped, first half of 2025 (SAG)Humanoids shipped, first half of 2026 (SAG)about 19,100010,00020,000humanoids
Crunchbase counts all robotics startups, and 2026 runs to 22 June. IFR counts full-size humanoids sold; SAG counts humanoids shipped.
How this was made
The funding bars are the yearly totals in Crunchbase News's chart of 22 June 2026; the 2026 bar is the year to that date. The humanoid counts come from two trackers with different rules: the IFR counts full-size humanoids, above 140 cm, sold in 2025, and SAG counts every humanoid shipped, by half-year. The panels share no scale and no ratio joins them, because the funding goes to every kind of robotics startup.

Venture funding raised by robotics startups, by Crunchbase's count, in billions of US dollars: 5.8 in 2020, 14.1 in 2021, 9 in 2022, 7 in 2023, 8.4 in 2024, a record 15 in 2025, then 18.8 in 2026 to 22 June. Humanoids counted: about 7,000 full-size humanoids sold in 2025 (IFR), and about 5,100 shipped in the first half of 2025 and 19,100 in the first half of 2026 (SAG). The funding covers every robotics startup, not only humanoid makers.

Funding by year and the humanoid counts
series and periodvaluecounted by
Robotics startup funding, 2020$5.8 billionCrunchbase
Robotics startup funding, 2021$14.1 billionCrunchbase
Robotics startup funding, 2022$9 billionCrunchbase
Robotics startup funding, 2023$7 billionCrunchbase
Robotics startup funding, 2024$8.4 billionCrunchbase
Robotics startup funding, 2025$15 billionCrunchbase
Robotics startup funding, 2026 to 22 June$18.8 billionCrunchbase
Full-size humanoids sold, 2025about 7,000IFR
Humanoids shipped, first half of 2025about 5,100SAG
Humanoids shipped, first half of 2026about 19,100SAG
Azevedo 2026, International Federation of Robotics 2026, Sui 2026.

Evidence: on September 21, 2026, Reuters reported that regulators were slowing humanoid listings after Unitree's shares rose more than fivefold on debut, then fell 55% from their peak, and quoted a venture capitalist saying some private-market projects had taken valuation cuts of 30% to 50%Source 27.

Three hypothetical failures could start a winter

The three mechanisms below are hypothetical scenarios. Their precedent is the 1973 Lighthill report, which found progress in robot-building "even more disappointing" than in the rest of AISource 28.

Speculation: the reliability gap outlasts investor patience. If 99.9% takes five more years of per-task engineering, the 2026 fund vintage marks to market against pilots. Bessemer says the last stretch "takes longer than most timelines assume"Source 6.

Speculation: the data thesis plateaus. If 10x more data stops buying success-rate gains, the fix is new science such as tactile integration, contact-faithful simulation or world models (data bottleneck). Ken Goldberg's 100,000-year data gap is the timing version: he accepts that data-driven methods "will eventually produce general purpose robots", and asks whenSource 10.

Speculation: the form factor is wrong. If task-specific systems keep winning cost per task, the humanoid thesis most of the capital is priced on collapses, as Brooks's "pure fantasy thinking" line predictsSource 1.

As of September 2026, none of the eight milestones is met

The board reads three not met, five partial, zero met. In unseen homes, π0.5 cleaned kitchens and bedroomsSource 11, and in September 2026 Figure reported 56% zero-shot success in 30 unseen homes, against 9% without Index pretrainingSource 29. Generalist reports 99% average success where earlier models reached 64%, though "not all tasks" reach those ratesSource 30. RoboArena, RoboChallenge and ManipulationNet are building shared evaluation infrastructureSource 14 Source 15 Source 16.

Figure's Helix 02 takes fingertip tactile inputSource 31, and T-Rex beats its strongest baseline by over 30%, though most VLA models still "overlook the tactile modality"Source 32. Sparsh-X and TouchWorld are tactile foundation modelsSource 17 Source 18. Surveys name contact modeling a key contributor to the sim-to-real gap, "particularly problematic in contact-rich tasks like robotic manipulation"Source 19.

Four measurable results would falsify the bear case

What would prove the bears wrong
Four results together would falsify the bear case
Suppose this result arrives
The plug turns only when all four pins sit on the lit lineNone of 4 in, 2 partly · locked

Pin status from the watchlist, September 2026; pin heights illustrative

Four results would each break one premise of the bear case. The lock opens only with all four; by September 2026, none had arrived.
How this was made

The bear case rests on four premises, and the article names the result that would break each: a verified deployment of 10,000 humanoids doing documented work (commercial viability), one policy above 90% across more than ten unseen homes (generalization), a cost-per-task parity analysis (form factor), and a real-world success-rate scaling law across tasks (the plateau). Each pin starts where the bear-case watchlist above puts that result in September 2026: not met for the deployment and the cost analysis, partial for the unseen homes and the scaling law. None is met.

The drawing is a pin-tumbler lock cut open in front of its pins, one stack per premise. A plug turns only when the split between every key pin (dark) and its driver pin sits on the shear line, the plug's top edge (lit). A result not yet shown leaves its key pin standing well above the line; partial evidence leaves it closer; supposing the result arrives sets it on the line. With all four on the line the plug turns, the key turns with it, and the bolt draws back out of its keeper.

Illustrative: the statuses are the watchlist's calls, but how high a pin stands only shows the call, not how far the evidence is from the result. The lock does not weigh the premises against each other; it shows that the four results are needed together.

Results from the article's final section; status calls from the bear-case watchlist (September 2026).

A verified deployment of 10,000 humanoids doing documented work breaks the commercial-viability premise, which vendor records so far measure in hoursSource 4 Source 5, and one policy above 90% across more than ten unseen homes breaks the generalization premiseSource 29 (generalization). Cost-per-task parity breaks the form-factor premiseSource 1.

A real-world success-rate scaling law across tasks breaks the plateau premise: EgoScale's scaling law is fit to validation lossSource 20, and Lin et al. find a "roughly power-law relationship" for single-task generalizationSource 33. Until then, more than $23B had been raised by early June against that scoreboard.