Lines, listings and design locks arrived, but deliveries did not
Read together, the fortnight points one way: production is starting and capital markets are opening, while customer deliveries, independent evidence and comparable specs are still a year or more away. Five conclusions, each with its status as of Nov 3, 2026.
Musk said on the Oct 21 Q3 call that Optimus production had begun at Fremont, that 20,000 robots a week is "at least a year" away, and that outside deliveries come around the end of 2027 Secondary summary. XPeng said on Oct 28 that IRON's solid-state battery is in small-batch production while the robot is still pre-SOP, with mass production planned for end-2026 and deliveries in H1 2027 Company plan. Agility alone shows delivered units: Digit has moved over 100,000 totes in more than 65,000 operating hours at GXO, while its 10,000-a-year RoboFab is a design capacity, not output. Module 4
Unitree's first report since its Aug 19 listing, filed Oct 28, shows Q3 revenue of CNY 665M (+69.93%) but net profit of 113M (-17.37%) Filed. Nine-month revenue was 1.817B and net profit 387M. Inventory reached CNY 1.019B against 368M at end-2025 while operating cash flow fell 54.69% to 194M, so the balance sheet says more about production than the income statement. No 2026 unit count is disclosed. Module 4
Churchill Capital Corp XI shareholders approved the Agility merger on Nov 2, with 21,069,042 votes for and 1,334,087 against Approved. The roughly $465M gross raise (164 + 201 + 100) is still "expected", with closing due Nov 4 and AGLT trading on Nov 5, so both are outcomes to verify Expected. Physical Intelligence's talks at about $20B and Generalist's reported $3B round carry no closing evidence. This is not investment advice. Module 5
Figure's Oct 22 all-hands disclosed Figure 04 in design lock with ten battery packs brought up and a fourth-iteration hand, but no price, date or public view of the robot. The Helix 3.0 teaser has no evaluation or release date, and Helix 2.5's 237 of 420 trials (56.4%) is a company-reported figure Company-reported. The only outside measurement in the model chapter is a latency test: GR00T N1.7 at 109 ms per call on Jetson Thor, from an inference vendor. Modules 2 and 3
The Information reported on Oct 31 that Meta built the Max prototype with Fourier in about three months (about 1.5 m, 48 kg, Jetson Thor); Meta calls it internal research with no plan to sell Reported. NVIDIA's GR00T 1.7 is open under Apache 2.0, and Agility's Digit 5 will use IGX Thor. Four of the platforms compared reference NVIDIA compute, so differentiation is moving to data, hands and batteries. Module 6
Capacity, listings and design locks arrived on schedule, but units delivered to paying customers beyond Agility's, and independent proof of model claims, remain 2027 events.
What changed at a glance
No new humanoid reached customers: five items, five evidence types
Between Oct 23 and Nov 2, 2026 the window brought a design-lock update, a press report of a prototype, a pre-order launch with 2027 delivery, and a factory claim from secondary sources.
Dated product table
Item by item
A status update, not a reveal
CEO Brett Adcock published a recording of an internal all-hands on Oct 22, one day before the window. Humanoids Daily says the design lock was first announced in May, so Oct 22 is a progress report on a known milestone. Status: design lock; components in testing.
What was said
Actuators were running in the test lab and ten full battery packs had been brought up the previous week, with load and thermal testing next. The hand was on its fourth iteration; Adcock called it approaching human parity, but no dexterity benchmark was shown.
Not yet seen
Figure built its 1,000th Figure 03 over the summer and cut cost by more than 90% from Figure 02 to 03, with no absolute cost or price given. Figure 04 was blurred on the generations slide. Humanoids Daily's own estimate for a reveal is late 2026 or early 2027 (outlet estimate, not Figure's).
A prototype known only through reporting
On Saturday Oct 31, The Information reported that Meta and Fourier built a humanoid prototype, Max, in about three months. Reuters relayed it but said it had not verified it independently.
Reported details
About 1.5 m and 48 kg (106 lb), NVIDIA Jetson Thor compute, five-fingered hands and a face display. Meta's robotics product vice president spent almost two months at Fourier's Shanghai factory. Max 2 has been in development since September (reported).
Meta's position
A Meta spokesperson said the work is internal research and experimentation, with no plan to sell the units. Reported targets such as recognising voices at 5 m are requirements, not verified capability; no independent evidence shows Max doing chores.
The only product with a stated price
Realbotix announced Echo on Oct 26; Onconetix summarised it on Oct 27. Status: launched, reservations open, nothing delivered. A $199 deposit reserves a unit through Realbotix's own site. The robots target companionship, education and wellness.
Onconetix's acquisition of Realbotix LLC, agreed Feb 12, 2026, is pending approvals. This is not investment advice.
A weaker claim
According to ABL.news and an X post, the Shenzhen company says it is "mass-producing and shipping" the wheeled humanoid and started a line "targeting tens of thousands of units" in H2 2026; its first factory was built for about 1,000 robots a year.
Label: company claim, unverified
Named uses include an HKC three-year agreement for more than 1,000 robots and Dongfeng Liuzhou material handling. No primary source, shipped-unit count, price or date is in the material. A line target is not units delivered.
Price and delivery date appear together only for Echo, a niche companion robot whose first units arrive in H2 2027. For Figure 04, Max and Optimus V3 the window supplied status and expectations, not products: read them as roadmap signals.
No robot foundation model has independent proof of its headline claim
Every success rate below is company-reported or second-hand. The only outside measurement in the material is a latency test. Events before Oct 23, 2026 are context.
Who claims what
Per-call latency on Jetson AGX Thor
ms per call, lower is faster. Third-party test by ZETIC, an inference-optimisation vendor that discloses its interest (BF16, batch 1, PyTorch eager; date not stated). Call times, not task success.
Model by model
56% in strangers' homes
Helix 2.5 (Sept 2026, context) reports 237 of 420 full-task trials (56.4%) across bed making, towel folding and toy tidying in 30 Bay Area homes it had never seen. The same policy trained from scratch scored 9%. No outside lab has repeated the test, so the 44% failure rate is the number to watch Company-reported.
A teaser, not a release
On Oct 22 Adcock teased Helix 3.0: more data and GPUs than any earlier Figure model, a larger one due "within weeks". He called an internal result "baby AGI" but gave no evaluation detail or release date.
Index data program
37.5 million uploads from 112 countries, 1.26 million downloads and an estimated $8.1 million data spend in the previous week (company-reported, Oct 22). The August launch figure was 16 million videos. Uploads are not cleaned training footage.
Open weights, N2 unconfirmed
The NVIDIA item dated in the window is a tutorial: on Oct 23 the developer blog walked through a humanoid workflow built on GR00T 1.7. NVIDIA calls it the first open, commercially usable VLA for general humanoid skills: Apache 2.0, a 3B-parameter checkpoint. No release date is given, so N1.7 is context.
Data and benchmark
Pretrained on about 32,000 hours of real demonstration and first-person human data plus about 8,000 hours of simulation (40,000 in total). Against N1.6 NVIDIA reports DROID-F6 +61% and DROID-F0 +10%, units not stated Vendor claim.
The one outside number
ZETIC measured 109 ms per call at 65.5 W and 7.11 J, against 254 ms for π0.5 eager. The material has no page on GR00T N2, so we make no claim about it.
Gemini Robotics
A video review of Oct 31 says Gemini Robotics ER 2 is a planner in preview in Google AI Studio, and the Gemini Robotics 2 motion model is trusted-testers only. TechTimes calls ER 2 an on-device motor-control model; we follow the narrower planner reading until a DeepMind page is available.
Physical Intelligence
Opened its Pi API beta on Oct 13 (context): waitlist only, no public price. Reported talks to raise at least $1 billion at about $20 billion are unconfirmed (module 5).
Other APIs
Dream Labs opened a public API on Oct 19 (context) at $3 an hour, each call returning 1.6 s of motion: 3,600 ÷ 1.6 = 2,250 calls an hour, about $0.0013 a call. PokeBot announced RoboAPI on Oct 29, a non-commercial API for Poke-Aurora that it says needs no task-specific training.
Second-hand figures
OpenAI's GPT-6 Astra is a general model with no OpenAI primary page in the material. 19 of 20 block trials (95%, RoboCurve, Sep 4) and 74.8% on 120 simulated tasks in NVIDIA's Astranomy study, at about $36.40 per successful episode and 18.56 s per direct call.
Galbot and the skeptics
Galbot reportedly lifted 10 RoboTwin tasks from 26% to 48% using Astra to supervise π0.5. Skeptics call the tasks simple.
Latency is the gap
18.56 s ÷ 0.109 s ≈ 170x GR00T's call time (different tests, order of magnitude only). Judgement: general models are taking planning, while fast embodiment-specific control stays specialist.
Figure's 56%, NVIDIA's benchmark gains and Astra's lab trials are vendor or second-hand figures; no robot foundation model in this window has independent proof of its headline claim.
Only Agility documents paid, delivered robots; the rest are lines and capacity targets
Status words are strict: a line installed or a capacity target is not units built, and units built is not units delivered to a paying customer. Galbot has no production or delivery figure in the material.
Production and delivery status at Nov 3, 2026
Unitree: growth with a profit squeeze
Change as reported, %. Filed Oct 28; BigGo Finance summary of the filing (CNY). Q3: revenue 665M, net profit 113M, excl. non-recurring 105M, R&D 127M. Nine months: revenue 1.817B, net profit 387M, excl. non-recurring 349M, R&D 263M.
The detail behind each maker
Early production, no deliveries
Musk said on the Oct 21 Q3 call that Optimus production had begun at Fremont and that scaling to 20,000 robots a week would take "at least a year, maybe a little longer". Tesla's own deck and transcript are not in the material, so this is a secondary summary. 20,000 × 52 ≈ 1.04 million a year, a design figure, not output.
A disputed reservation date
On Oct 30 The San Francisco Standard reported, from a source, that Tesla would open Optimus reservations in early December, possibly around the Dec 1–2 Humanoids Summit, with a price above $20,000 Reported, unconfirmed. Musk replied "No" on X the same day, without saying which part he denied. Tesla has confirmed no price or deposit.
First report since listing
Unitree (STAR Market, 688836) reported Q3 on Oct 28, its first periodic report since its Aug 19 listing. Q3 R&D is 127 ÷ 665 = 19.1% of revenue; Q3 profit is down about 49.6% from Q2's 224M (our arithmetic). Unitree attributes the nine-month profit jump to gross profit growth and a one-off share-based payment expense last year.
The balance sheet says more
Inventory reached CNY 1.019B versus 368M at end-2025 (+177%, about 2.8x), and nine-month operating cash flow fell 54.69% to 194M. No 2026 unit count is in the material. The stock closed at 695.90 yuan on Oct 28 (market cap CNY 281.5B); not investment advice.
A component milestone
Vice president Gu Jie said at SAECCE on Oct 28 that IRON's solid-state battery is in small-batch production while the robot is still before SOP Company statement. No output, capacity or energy-density number was given, and XPeng did not say whether the cells are all-solid-state or semi-solid.
Plan, not volume
The Sept 8 line commissioning (over 80% of core processes automated) was a demonstration. Plan: mass production end-2026; deliveries in H1 2027 per Pandaily and AUTONOMOUS, while CnEVPost says only "2027". We follow the more specific H1 2027 as a company plan.
Delivered, and growing by site
On Oct 26 Agility and GXO announced Digit at two more GXO sites, Statesville (NC) and Lebanon (TN), deployments expected from early 2027 on Digit 4. At Flowery Branch (GA) Digit has moved over 100,000 totes and the fleet has over 65,000 operating hours Company release.
Potential, not revenue
Three unnamed manufacturers are in paid programs; AGCO's Hesston pilot (announced Oct 20, before the window) starts in 2027. The $300M+ Digit 5 orders are potential multi-year value, not revenue. RoboFab's 10,000 a year is a design capacity.
AI² Robotics AlphaBot 2
Says it is mass-producing and shipping AlphaBot 2, with HKC's three-year agreement covering 1,000+ robots (signed, not delivered) and a new line targeting tens of thousands, versus about 1,000 a year at its first factory Company claim.
AgiBot
AgiBot's Oct 29 Innovation Day launched the AHA3 and a "Worken" metric meant to replace unit shipments. IDC and SAG reportedly show its H1 shipments passing Unitree's, but we found no primary AgiBot unit figure.
Only one deal is filed; most other money is reported or undated
In the window, the one closed, filed event was Agility's SPAC vote. Agility's shareholders approved the Churchill XI merger on Nov 2, but the $465 million raise is still "expected": no closing filing is in our sources.
Funding and deal table
Sources: SEC 8-K (vote); Agility release (raise); Bloomberg via Stock Analysis (Physical Intelligence); trackers and press (others). Apptronik: about $935 million raised in total; valuations are unconfirmed secondary figures.
Agility's expected gross raise
$M, before transaction expenses. Trust account (after the redemption deadline), PIPE led by Foxconn, and equity already received after the deal announcement. Expected, per the Oct 30 company release: 164 + 201 + 100 = $465M. Secondary sources cite a $2.5B pre-money value against about $1.8M of 2025 sales, about 1,400x. Not investment advice.
The Nov 2 vote and what is still to verify
Approved, with a caveat
About 84% of the 25,026,951 shares present voted for (about 44.93% of voting power attended). The 13.8 million Class B shares are in the total and the 8-K does not split votes by class, so the tally alone does not show how public holders voted Approved.
Outcomes to verify
Closing is expected Nov 4 and Nasdaq trading as "AGLT" on Nov 5, both upcoming as of Nov 3; existing holders roll equity and accept a 180-day lock-up. The 8-K reports approval, not closing or the final redemption count Expected.
Private rounds, talks and partnerships
Physical IntelligenceIn talks
Bloomberg reported early talks to raise at least $1B at roughly $20B, with Avenir and Lux Capital in discussions to lead. This is "in talks", not a round, and the date is not stated. Its last closed round was about $1B at above $11B led by Founders Fund ($11.19B at a May 21, 2026 valuation date). The company has no revenue.
Generalist AIReported closed
TechCrunch, via Intelligent Living, reports a Series B extension of nearly $200M led by 8VC at $3B, six weeks after a $400M round at $2B; about $600M for the round, backed by a Form D. Date not stated.
ApptronikClosed (context)
The $520M extension in February 2026 raised its Series A to over $935M. Valuation is $5.30B per Eqvor; another source says $5.5B. We use about $935M and label valuations unconfirmed.
China: ROBOTERA, Galbot and othersMixed
ROBOTERA's May 8 round (over $200M, led by SF Group) is pre-window. Galbot's $100M round has no date in Dealroom's page. Trackers disagree on China totals (about $1.15B company-reported versus about $980M itemised); we follow the about $980M itemised figure as the documented floor. Bleu Robotics: EUR 5M for software that teaches industrial robots tasks from fewer than ten demonstrations.
Commercial deals and partnersSigned / reported
GXO signed to add Statesville, North Carolina and Lebanon, Tennessee, starting early 2027 with Digit 4; no contract value is given. Meta's Max prototype with Fourier is a reported collaboration, internal research. NVIDIA appears as an Agility investor and as supplier of Jetson Thor in Max and of IGX Thor in Digit 5 (announced Oct 22, context).
Routes differ on legs, hands, batteries and data, while compute converges on NVIDIA
Specs are company-reported, mostly via secondary summaries dated July to October 2026, so they are context rather than window events. "Not disclosed" marks wherever a maker has published nothing.
Comparison table
Route trade-offs
What the sources say
A motor supplier's guide (Mar 25, 2026) rates wheeled bases lower in cost and maintenance and faster on flat floors, legs higher in terrain reach and complexity. Asian Robotics Review argues flat factories give legs no advantage; that is opinion, and its "nearly three times faster" claim is unsourced.
The split is not clean
Wheeled Galbot showed a biped ET1 in August 2026 (context), and Apptronik's Apollo 2 offers a bipedal or wheeled base. AI² Robotics' evidence is deployment claims (HKC 1,000+ robots, company claim). No independent head-to-head test is in the material.
Hands are a data decision
Atlas publishes 56 DoF; NEO's sheet lists hands 22 each, arms 7, legs 6, spine 2, neck 3, so 2×22 + 2×7 + 2×6 + 2 + 3 = 75 DoF, with hands 44 ÷ 75 = 59% of the total (our arithmetic). The GR00T reference robot's DoF figure is not in the material.
Figure discloses no DoF
Adcock said in May that limited degrees of freedom "pollute" training data, and the fourth-iteration Figure 04 hand is described as approaching human parity without a benchmark. The hand is the interface to human video (Index).
Runtime is not productive time
Atlas swaps batteries and Figure 03 charges wirelessly; runtimes are battery duration. Digit 5's 9-minute charge per 90 minutes is 9 ÷ 90 = 10% of runtime.
Not yet comparable
XPeng said on Oct 28 that IRON's solid-state battery is in small-batch production, chemistry unstated and robot still pre-SOP; Figure 04's batteries are still before load and thermal tests. Neither discloses energy density.
GR00T 1.7 data mix
About 32,000 h of real and egocentric human data plus about 8,000 h of simulation: 32,000 + 8,000 = 40,000 h, 80% real or human (NVIDIA blog dated Oct 23, release date not stated). NVIDIA's workflow adds teleoperation collection.
Human supervision
1X leans on scheduled human supervision (Expert Mode), and no primary source gives its autonomy share.
Jetson Thor and IGX Thor
Reported Max prototype and Galbot ET1 use Jetson Thor; Digit 5 will standardise on IGX Thor (announced Oct 22, one day before the window). NVIDIA cites 25 Hz for GR00T 1.7 with TensorRT NVFP4.
A third-party number
ZETIC, an inference vendor, measured 109 ms per call in eager PyTorch (1,000 ÷ 109 ≈ 9 Hz) at 65.5 W; 65.5 W × 0.109 s = 7.1 J matches its 7.11 J. Date not stated.
Vertical integration is only partly true
Tesla Fremont lines were being installed per its July 22 update, with initial units feeding Optimus Academy data collection. Figure owns hardware, Helix and Index data, yet calls parts of Figure 03 "hard to manufacture" and plans die-casting for Figure 04. XPeng's IRON still relies on external actuator, dexterous-hand, sensor and AI-chip suppliers.
Reference platforms are NVIDIA's route
Open GR00T models, Jetson Thor, Halos safety and the Agile ONE SN reference robot. An NVIDIA blog (date not stated) lists Humanoid, LG Electronics, NEURA and Noble Machines as GR00T adopters, and Agility and Boston Dynamics as Cosmos and Isaac users.
Partner builds split body and brain
Reportedly, Meta's Max (1.5 m, 48 kg) was built with Fourier in about three months, with Max 2 started in September; Meta says it is internal research. Google's Gemini models paired with Apptronik's Apollo (Tech Times, Oct 23) are the model-plus-partner version.
Our judgement: four platforms (Max, Galbot ET1, Digit 5, GR00T) reference NVIDIA compute, so dependence is concentrating at silicon and base models, and differentiation is moving to data, hands and batteries. The specs that would settle the comparison for Figure 04, Optimus and IRON are undisclosed.
Capacity, listings and design locks are in; customer deliveries and outside proof are 2027 events
The route differences are real but undisclosed in the specs that matter. This module turns the window into implications for three readers and a dated watch-list; dates already past on Nov 3, 2026 are marked by their outcome, not as to-dos.
What changed and what it means, by reader
What changed
Tesla and XPeng started lines but deliver in 2027; Agility reports 100,000+ totes and 65,000+ operating hours at GXO and adds two sites from early 2027.
Implication
Separate pilot from production: ask for units delivered and operating hours, not line capacity. Only Agility documents paid, delivered robots; Digit 4 is the only humanoid here with documented paid deliveries. Digit 5 early access is H1 2027, general availability end-2027.
Our reading
The line, not the robot, is what shipped. Tesla's 20,000-a-week target would be a 40x step from an illustrative 500 a week (our illustration), and no unit count exists.
What changed
Unitree's first report: Q3 profit down 17.37% while inventory rose to CNY 1.019B; Agility's merger vote passed Nov 2, closing expected Nov 4.
Implication
Listing routes are open, but valuation runs ahead of revenue: $2.5B against about $1.8M of 2025 sales is about 1,400x (secondary figures). Not investment advice.
Our reading
The STAR Market gave a first quarterly read; the US SPAC route cleared its vote with 21,069,042 of 25,026,951 present shares (about 84%) but the final redemption count and closing are not yet in a primary source. Reported rounds such as Physical Intelligence's talks are not closed rounds.
What changed
GR00T 1.7 is open (Apache 2.0); Figure's Helix stays closed with a 56.4% company-reported home result; Astra's figures are second-hand.
Implication
Open stacks give a measurable baseline (109 ms per call on Jetson Thor, third-party); closed models give claims without outside tests.
Our reading
Fast control is specialist: 109 ms per call for GR00T against 18.56 s for a direct Astra call, though the tests differ and give only an order of magnitude. Dependence is concentrating at silicon, since Max, Galbot ET1, Digit 5 and GR00T all reference NVIDIA compute.
What would change this reading
1 A primary filing
A filing or release that confirms the Agility closing and the final redemption figure after the 8-K.
2 A Tesla statement
A statement, rather than a source-based report, on reservations, price or unit counts; Musk's one-word reply did not say which part he denied.
3 An outside test
An independent test of a headline model result, since Helix 2.5's 237 of 420 and Astra's 74.8% remain company or second-hand figures.
GR00T N2 and the Agile ONE SN reference robot have no date in our sources, so they are not on the watch-list.
Key dates
Evidence moved from demos to documents this fortnight: filings, commissioned lines and signed sites. Yet those documents describe capacity, plans and votes, and customer deliveries beyond Agility's remain 2027 events.