The Bear Case

Every other module in this domain documents genuine progress, and the progress is real. This module gives the opposite reading the same evidentiary standard. The bear case says robotics is repeating the pattern that produced earlier AI winters: capital arrives on demonstrations, deployment evidence arrives late or never, and the funding window closes before the science closes the reliability gap. What follows is that case stated by its strongest named proponents. Every item is labeled evidence (a published position, a filing, a documented event) or speculation (a scenario about what happens next), and the module ends with the eight milestones whose outcomes would settle the argument either way.

Here is the scoreboard the argument turns on. Eight milestones, each chosen so its outcome moves the bear case whoever is right about everything else. Switch the status filter to Met and the board empties; switch it to Partial and four rows survive, each with the published evidence behind the call.

The skeptic record

Evidence: Rodney Brooks, co-founder of iRobot and Rethink Robotics and now running Robust.AI, is the field's most prominent named skeptic, and his claim is specific. His September 2025 essay calls the belief that current methods will produce practical humanoid dexterity within decades "pure fantasy thinking," on tactile grounds: vision-only training gives the robot no force channel, and Johansson's anesthetized-finger experiments show what humans become without one. Dexterity covers the argument in full Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. His January 2026 predictions scorecard holds the line: deployable dexterity "will remain pathetic compared to human hands" beyond 2036, and walking humanoids stay too dangerous near people without new mechanical systems Brooks 2026↓Predictions Scorecard, 2026 January 01Rodney Brooks, 2026. These are dated, falsifiable predictions from someone who has shipped robots for four decades, which is what separates them from ambient doubt.

Evidence: Morgan Stanley's July 2026 note argues the industry has a "PR problem": humanoids are publicly framed as direct substitutes for workers rather than as tools for hazardous or labor-constrained tasks, so "the industry's social license to deploy may matter just as much as technical performance" Wilkins 2026↓'PR problem' is standing in the way of China's humanoid robot boom, says Morgan StanleyJoseph Wilkins, CNBC, 2026. The same note says investors "are increasingly looking for tangible evidence of real-world return on investment" rather than polished videos, and flags the US ban on Chinese humanoid imports as an R&D cost increase. The bank kept its 50,000-unit China shipment estimate for 2026. This is not a short call; it is a bank telling clients the bottleneck has moved from capability to proof.

Unit economics

Evidence: Unitree, the volume leader, shipped roughly 5,500 humanoids in 2025, more than any Western competitor Noreika 2026↓Humanoid Robots in 2026: What Is Actually DeployedAlius Noreika, 2026, and then saw Q1 2026 adjusted net profit fall 52.55% year over year Ramsey 2026↓Robot Boom Meets Earnings Reality: Unitree Profits Halved, Optimus Not for SaleMireya Ramsey, TechTimes, 2026. Volume and profit are different variables, and the bear case starts there: the company winning the shipment race is not yet winning the income statement.

Evidence: Figure's private valuation of $39B (September 2025) exceeds Goldman Sachs' projection for the entire humanoid market in 2035, $38B Noreika 2026↓Humanoid Robots in 2026: What Is Actually DeployedAlius Noreika, 2026. One private company is priced above the projected size of its whole market nine years out, so the valuation is a bet on the market outgrowing the projection, not on current revenue.

Evidence: the inputs are expensive too. Bessemer estimates more than $3B of aggregate robot data costs over the next two years and frames that spend as the moat Levine 2026↓Bessemer Predicts: Robotics and physical AIJeremy Levine, Talia Goldberg, Janelle Teng Wade et al., 2026. A moat made of capital is also an exposure: the spending requirement that protects incumbents is the same burn rate the funding cycle has to keep feeding. What a shipped cell actually has to earn back, against integrator margins and jam-clearing labour rather than against research budgets, is priced in Industrial Deployment.

Evidence: at Computex 2026, a Qualcomm-powered humanoid collapsed face-first during the live keynote, a failure attributed to a communication glitch, and was covered and carried off stage Malayil 2026↓Watch: Qualcomm-powered humanoid robot collapses during live keynoteJijo Malayil, Interesting Engineering, 2026. One fall says nothing about any single program. As a public datum about the distance between stage demos and floor reliability, it is on the record.

The demo-to-deployment gap

The Reliability Gap documents the arithmetic: 80% per step is a demo, 99.9% is a product, and compounding makes the distance between them brutal at thirty steps. The bear case is that closing the gap takes longer than the capital lasts. Lisa Yan, quoted in Bessemer's own analysis, describes the climb from 99% to 99.9% as steep Levine 2026↓Bessemer Predicts: Robotics and physical AIJeremy Levine, Talia Goldberg, Janelle Teng Wade et al., 2026. Morgan Stanley's ROI warning is the market making the same point from the other side Wilkins 2026↓'PR problem' is standing in the way of China's humanoid robot boom, says Morgan StanleyJoseph Wilkins, CNBC, 2026. And the strongest verified deployment records, Agility's 65,000+ operating hours and Figure's 1,250+ hours at BMW Spartanburg, are supervised programs in narrow applications, which is what a bridge looks like before it becomes autonomy Noreika 2026↓Humanoid Robots in 2026: What Is Actually DeployedAlius Noreika, 2026. The evaluation crisis compounds the problem: when every demo is its own benchmark, even genuine progress is hard to underwrite.

The capital cycle

Evidence: robotics startups raised more than $23B globally by early June 2026 by PitchBook's tally, closing in on the $26B raised in all of 2025 Savage 2026↓Robotics Startups Raised $23 Billion in 2026, Closing In On All of 2025Andre Savage, Market Briefs, 2026. Crunchbase's narrower venture count puts the figure at $18.8B, already past the $14.1B peak set in 2021 Azevedo 2026↓Sector Snapshot: Robotics Startups On Fire As Venture Funding Surges To Record Numbers In 2026Mary Ann Azevedo, Crunchbase News, 2026. The two tallies disagree about the number and agree about the shape: capital is arriving faster than deployment evidence.

Whether that inflow constitutes a bubble is itself contested, and both sides are named. Bessemer Venture Partners argues there is "no robotics bubble" and that the sector is "structurally underinvested," while conceding that "not every company being funded will succeed" and that "some valuations are stretched" Levine 2026↓Bessemer Predicts: Robotics and physical AIJeremy Levine, Talia Goldberg, Janelle Teng Wade et al., 2026. The skeptic reading, Brooks's timeline predictions and Morgan Stanley's demand for ROI evidence, takes the same inflow as too much capital chasing demonstrations Brooks 2026↓Predictions Scorecard, 2026 January 01Rodney Brooks, 2026 Wilkins 2026↓'PR problem' is standing in the way of China's humanoid robot boom, says Morgan StanleyJoseph Wilkins, CNBC, 2026. This module does not resolve the disagreement; Competing Theses maps the technical version of it. The winter scenario needs only the disagreement itself: the market has not converged on how to value these companies, and unconverged valuations reprice violently.

How a winter would happen

The mechanisms below are scenarios, not predictions. Each rests on the evidence above, and each is labeled accordingly.

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 rather than products, and the pullback slows every program, including the ones whose science is working.

Speculation: the data thesis plateaus. If 10x more data stops buying success-ratesuccess rateThe standard headline metric of robot learning evaluation: the fraction of attempted episodes in which the policy completes the task. Trial count, time limit and scene distribution belong with the reported rate. Toyota Research Institute's Large Behavior Model study analyzed 1,800 real-world rollouts and warned of a significant risk of statistical noise from insufficient power, not a measured prevalence of incorrect papers. An inconclusive comparison does not establish policy equivalence. gains, the fix is new science (tactile integration, contact-faithful simulation, world models), and science does not arrive on a fund's schedule. Ken Goldberg's 100,000-year data gap is the strongest statement of this premise Goldberg 2025↓Good old-fashioned engineering can close the 100,000-year "data gap" in roboticsKen Goldberg, Science Robotics, 2025.

Speculation: the form factor is wrong. If task-specific systems keep winning cost per task, the humanoid thesis that most of the capital is priced on collapses and takes the funding cycle with it, even though warehouse automation keeps working. Brooks's "pure fantasy thinking" line is the named version of this scenario Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025.

The watchlist

Back to the board at the top of this module. Each row carries the question it settles, a current-status call with the published evidence behind it, and the observation that would flip it to met. As of August 2026 it reads four not met, four partial, zero met.

Several rows are closer than the labels suggest. π0.5 has already cleaned kitchens in homes it never trained on Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025, RL-100 and π*0.6 hold the per-task reliability records Lei 2025↓RL-100: Performant Robotic Manipulation with Real-World Reinforcement LearningKun Lei, Huanyu Li, Dongjie Yu et al., 2025 Amin 2025↓π*0.6: a VLA That Learns From ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025, and RoboArena, RoboChallenge, and ManipulationNet are converging on shared evaluation infrastructure Atreya 2025↓RoboArena: Distributed Real-World Evaluation of Generalist Robot PoliciesPranav Atreya, Karl Pertsch, Tony Lee et al., 2025 Yakefu 2025↓RoboChallenge: Large-scale Real-robot Evaluation of Embodied PoliciesAdina Yakefu, Bin Xie, Chongyang Xu et al., 2025 Chen 2026↓ManipulationNet: An Infrastructure for Benchmarking Real-World Robot Manipulation with Physical Skill Challenges and Embodied Multimodal ReasoningYiting Chen, Kenneth Kimble, Edward H. Adelson et al., 2026. Others are early. Tactile foundation models such as Sparsh-X and TouchWorld exist but sit outside every major VLAvision-language-action modelA robot policy built by adapting a vision-language model, pretrained on web-scale image and text data, to output robot actions. RT-2 established the approach by representing actions as text tokens and co-fine-tuning on web and robot data together, showing that semantic knowledge from web pretraining transfers into control. The same recipe underlies OpenVLA and the π series of generalist policies. pipeline Higuera 2025↓Tactile Beyond Pixels: Multisensory Touch Representations for Robot ManipulationCarolina Higuera, Akash Sharma, Taosha Fan et al., 2025 Zhou 2026↓TouchWorld: A Predictive and Reactive Tactile Foundation Model for Dexterous ManipulationJianyi Zhou, Feiyang Hong, Yunhao Li et al., 2026, contact-rich sim-to-realsim-to-real gapThe discrepancy between a simulator's physics and rendering and the real world, which makes policies trained in simulation degrade when they are deployed on hardware. The gap has several independent components, dynamics, sensing, and actuation among them, and the standard transfer methods (domain randomization, system identification, teacher-student distillation, real-to-sim reconstruction) each attack a different part of it. remains the unsolved case in reality-gap surveys Aljalbout 2025↓The Reality Gap in Robotics: Challenges, Solutions, and Best PracticesElie Aljalbout, Jiaxu Xing, Angel Romero et al., Annual Review of Control, Robotics, and Autonomous Systems 2026 (accepted), 2025, and EgoScale's scaling lawscaling lawAn empirical regularity between a system's performance and the resources it consumes, parameters, data, or compute, first made precise for language models and now measured for robot learning. Lin and colleagues fit imitation-learning success to the training data and found that generalization to new objects and environments tracks data diversity rather than raw hours. EgoScale extended the measurement to egocentric human video, reporting a log-linear relationship between hours of human data and dexterous-manipulation success across four orders of magnitude. is measured on validation loss rather than real-world success rate Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026.

What would prove the bears wrong

The bear case is falsifiable, which is what makes it worth taking seriously. A verified 10,000-unit deployment breaks the commercial-viability premise. One policy above 90% across more than ten unseen homes breaks the generalization premise, and Generalization tracks how close the field is. A published cost-per-task parity analysis breaks the form-factor premise. A real-world success-rate scaling law breaks the plateau premise. Until those arrive, the honest summary is the board itself: zero of eight milestones met, four moving, and more than $23B raised by early June against that scoreboard.