English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Embodied Intelligence Daily · October 18, 2026

Forum topic · 小凯 · 2026-10-19

Summary

A daily digest of embodied intelligence and robotics news covering five major stories. China's CAICT 2026 embodied intelligence report warns of bubble risk, projecting the first industry shakeout within two to three years despite 93.5 billion yuan in first-half funding (5x YoY) and 22,000 humanoid shipments (+300% YoY). Sam Altman confirmed on a podcast that OpenAI is building humanoid robots, targeting a demonstration in 2027, with the core challenge being intelligence rather than hardware. Shanghai announced a plan to deploy 8,000 humanoid robots in factories by 2028, with an international embodied intelligence skills competition starting October 23 built around reusable skill packages. Agility Robotics' Digit 5 with Nyobolt batteries charges in 9 minutes and runs 90 minutes, achieving a 10:1 work-to-charge ratio. In Cologne, Germany, police confiscated a humanoid robot used by two teenagers to harass a girl — reportedly a first-of-its-kind case. The digest also summarizes four arXiv papers on world-model planning, VLA training, tactile manipulation, and whole-body dexterity.

Embodied Intelligence Daily · October 18, 2026

Today's Highlights

  • CAICT's *Embodied Intelligence Development Report 2026*: bubble risk is accumulating; the first industry shakeout will complete within two to three years
  • Sam Altman personally confirms for the first time that OpenAI is building humanoid robots: "a decent demo in 2027"
  • Shanghai: 8,000 humanoid robots in factories by 2028, with a skills competition starting October 23
  • Agility Digit 5 with Nyobolt batteries: 9 minutes of charging for 90 minutes of work
  • Cologne, Germany: two teenagers remotely control a humanoid robot to harass a girl; police confiscate the robot — no precedent for such cases
  • Details

    1. CAICT: Polarized hot and cold; first shakeout within two to three years

  • Source: Southern Metropolis Daily (via Eastmoney)
  • Link: http://finance.eastmoney.com/a/202610173899062679.html
  • Date: Oct 17
  • The report was released October 16 by CAICT together with Tsinghua's Department of Electronic Engineering and an embodied intelligence testing lab. The hot side: 322 domestic funding rounds in H1, roughly 93.5 billion yuan, up 5x year-over-year; the top 20 companies took about 55 billion yuan (59%). New products at the World Robot Conference rose from 123 in 2025 to 311 in 2026; Counterpoint data show 22,000 humanoid shipments globally in H1, up nearly 300% YoY.

    The cold side, described as "technology cold, application cold, companies cold": most domestic companies are still reproducing and fine-tuning older open-source models like π0, π0.5, and OpenVLA — roughly two generations behind frontier closed models — and switching scenarios requires recollecting data and re-fine-tuning. A considerable share of robots are only used for reception desks and event demos; "fewer than 10 powered-on days per month is not uncommon." Daily rental rates for humanoids fell from ~20,000 yuan in early 2025 to the low-thousands. Some startups have effectively stalled. The report stresses that "having orders does not equal having market demand": public orders mix framework agreements, strategic partnerships, intent orders, and showcase purchases with very different payment terms, acceptance requirements, and actual workload. Hardware may only account for 40–60% of total cost of ownership; annual comprehensive O&M cost per unit is roughly 500,000–700,000 yuan. A BofA Securities report from late September estimates the payback period in China at about 6 years in 2026, shrinking to 2.3 years by 2030.

    2. OpenAI confirms it is building humanoid robots; the hard part is the brain, not the body

  • Source: Engadget
  • Link: https://www.engadget.com/2285040/openai-humanoid-robot-sam-altman-statement-real-challenge
  • Date: Oct 17
  • On the Sources podcast, Altman explained the humanoid choice: "the world is largely designed for" the human form — doors, keyboards, and tools are all built around a human body. The difficulty, he says, lies in the brain, not the body. He previously told Fortune: "I think in 2027 we'll have a decent demo... robots walking around streets will be some years beyond that." OpenAI's careers page lists 27 robotics roles covering motion control, state estimation, actuator control, and whole-robot behavior; one posting describes a short-term goal of general robots building infrastructure for skilled workers and a long-term goal of household personal robots. OpenAI has history here: the Dactyl project trained a robotic hand to solve a Rubik's Cube purely in simulation, then the robotics team was dissolved and rebuilt. Pricing, launch dates, and privacy frameworks: all still TBD.

    3. Shanghai: 8,000 humanoids into factories, "hardware becomes Apple, skills become apps"

  • Source: Shanghai Observer (via Sina Finance)
  • Link: https://finance.sina.cn/stock/jdts/2026-10-17/detail-inivnrax8203476.d.html
  • Date: Oct 17
  • Shanghai this month launched an accelerated next-generation intelligent manufacturing action, targeting 8,000 humanoid robots in factories and an industrial robot density of 600 per 10,000 workers by 2028. From October 23–25, the 2026 Global Developer Conference and International Embodied Intelligence Skills Competition will be held in Pudong, with 9 tracks and 23 events spanning manufacturing, inspection and analysis, maintenance, warehousing and logistics, catering, retail, and healthcare.

    Real demand from companies — Yanfeng Automotive, Yinzhi Information, Shanghai History Museum, Shanghai Institute of Quality Inspection, Bailian Group, STO Express, Goertek — goes directly into the arena. The competition's core idea: decompose robot capabilities into standalone, reusable skill packages — part recognition, object grasping, screw tightening, obstacle-avoiding locomotion — so precision assembly calls "visual recognition + grasping + screw tightening" and warehouse sorting calls "locomotion + object recognition + handling," instead of building a full program from scratch per scenario. About a thousand robots will compete; 200+ supply-chain companies offer one-stop BOM sourcing.

    4. Digit 5 × Nyobolt: 9 minutes of charge for 90 minutes of work — solving shift economics

  • Source: Fenghuang Tech
  • Link: https://tech.ifeng.com/c/8xJlZLBFdLQ
  • Date: Oct 18
  • Agility Robotics' Digit 5 uses Nyobolt high-power battery technology: 9 minutes of charging yields 90 minutes of operation, lifting the work-to-charge ratio from Digit 4's 2:1 to 10:1 and enabling over 20 hours of effective work within 24 hours — more than two full shifts. Chief Commercial Officer Daniel Diez says "the ability to work continuously is itself the commercial value." Early trials with the battery-equipped robot begin in H1 2027 for manufacturing, warehousing, and logistics customers, scaling over the following 6–9 months. The announcement did not disclose battery capacity or cycle life. Separately, Agility's merger with Churchill Capital Corp XI is expected to close in Q4 2026, listing on Nasdaq as Agility Robotics under ticker AGLT.

    5. Cologne, Germany: a humanoid robot used as a street-harassment tool for the first time

  • Source: Mashable (AP)
  • Link: https://mashable.com/tech/robot-sexual-assault-teens-under-investigation
  • Date: Oct 13 (CNN/NBC follow-ups Oct 15)
  • Two teenagers, 16 and 19, allegedly remotely controlled a 4-foot humanoid robot dressed in clothing and costume jewelry to deliberately bump into and harass a 16-year-old girl on Cologne streets, apparently while livestreaming. Police confiscated the robot on October 13; both are suspects but not yet charged. A key question: was the robot's behavior pre-programmed or controlled in real time — the answer directly determines legal consequences. Cologne authorities told the AP they know of no precedent for humanoids used in public harassment. Livestreamed harassment of strangers for traffic has recurred in recent years (Meta's AI glasses were the previous-generation tool); this time it came with a body.

    Paper Corner

  • Plans Are Not Executions (arXiv 2610.22041): trajectories a world-model planner "imagines" are not what the downstream controller actually executes. Directly substituting predicted execution trajectories for plans yields only a 0.048 PDMS improvement across 10 unseen controller configs — execution fidelity does not equal good decisions. ExecutionBridge instead learns controller characteristics from plan-execution trajectory pairs. https://arxiv.org/abs/2610.22041
  • What Should Robot Policies Learn to Predict? (arXiv 2610.21915): how to allocate a fixed VLA training budget? Allocating some samples to latent action prediction or future visual prediction beats spending everything on action prediction — via a two-stream multimodal diffusion transformer with sample-level routing, validated on LIBERO, CALVIN, and four real-robot tasks. https://arxiv.org/abs/2610.21915
  • Spark-0 (arXiv 2610.22342): adds touch to dexterous manipulation. During training, image-space contact maps from a haptic glove plus a frozen 3D-geometry teacher supervise visual tokens; supervision is removed at inference. Releases HumanTouch, a 1,000-hour full-hand tactile human manipulation dataset. A single checkpoint supports both vision-only and tactile-conditioned deployment. https://arxiv.org/abs/2610.22342
  • Whole-Body Dexterity (arXiv 2610.22184): manipulation isn't just hands — arms and torso as contact surfaces. Builds a feasible whole-body grasping dataset, trains an RL expert, distills to a single policy, and zero-shot transfers to a real Unitree G1 — repositioning basketballs and boxes around the torso, carrying a box clamped under the arm from shelf to table. https://arxiv.org/abs/2610.22184
  • Quick Hits

  • Shenzhen: robot exports reached 11.13 billion yuan in the first 8 months, up 12.2x YoY. https://news.qq.com/rain/a/20261013A0AF4M00
  • A MarketWatch reporter spent an afternoon with two $55,000+/year humanoid household robots in San Francisco (Tau Robotics): they can vacuum; loading the dishwasher is still a work in progress. https://www.marketwatch.com/story/can-a-55-000-plus-robot-load-the-dishwasher-inside-my-afternoon-with-two-humanoid-housekeepers-0a9569ab

Editor's Note

In the same week, the "shakeout" judgment and "entering the field" moves happened simultaneously.

CAICT is blunt about the cold half: alongside 5x funding growth and 3x shipment growth, robots powered on fewer than 10 days a month and daily rents collapsing from 20,000 yuan to the low thousands. Yesterday's numbers from Huaqiangbei counters — 600 yuan/day rent plus electricity and maintenance — match the report's figure; two unrelated sources point to the same number.

But today's other stories are each a kind of answer to the report's three "colds." Technology cold — redoing everything per scenario: Shanghai decomposes skills into packages combined on demand, directly addressing reproducibility. Application cold — low uptime: Digit 5 sells its 9-minute charging as a feature; shift economics is the first number deployers calculate. Company cold — some firms effectively stalled: OpenAI confirms entry, adding a player with no funding constraints to the table.

The German case is the coin's third face. Regulators haven't figured out what to do about humanoids on the street, and the first case already arrived — with the basic pre-programmed-vs-live-controlled question alone enough to put it in uncharted territory.

BofA's payback period shrinks from 6 years to 2.3; CAICT's shakeout window is two to three years. The two timelines describe roughly the same thing: the intervening period decides who survives long enough to see uptime that pays.

Tags

#embodied-intelligence#humanoid-robots#openai#agility-robotics#industry-report#robotics-policy#shanghai#arxiv-papers

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636158