DossiersDosare · 16 Oct 202616 Oct 2026
Physical AI Tracker — verification, 2026-10-16
This report exists in English only.
title: Physical AI Tracker — verification pass created: 2026-10-16 owner: eth (code-eth) verifies: Research/physical-ai-tracker.md (as of 2026-10-16 09:00 entry) method: grep against eth-memory/sol/reports/.md, Research/ai-watch/.md + en/, Logs/model-watch.md. No web browsing — per house rule, a claim with no house source is UNSUPPORTED, not "checked online."
Checked every claim named in the ask, plus a spot-check pass across the entity tables (public + private) that those claims sit inside. Not a line-by-line re-derivation of all 132 lines — a targeted pass on the numbers and names most likely to travel wrong between a web source and a tracker row.
Priority items (named in the ask)
| Claim (tracker location) | Verdict | Evidence |
|---|---|---|
| Google Argon — row 2/7: "Video-to-Sim-to-Real; sim data lifted real success 56%→75%"; ERQA 87%; test bodies "Spot, Apollo, ALOHA 2, SO-101" | UNSUPPORTED | Gemini 4 Argon and Gemini 4 Argon for Robotics are real and well documented — GA Oct 13, confirmed in Logs/model-watch.md:8377-8390 ("Gemini 4 Argon went GA on October 13... Also launched: Gemini 4 Argon for Robotics"). But none of ERQA, "Video-to-Sim-to-Real," the 56%→75% figure, or the SO-101/ALOHA 2/Spot/Apollo body list appear anywhere in house files for Argon. Those bodies (ALOHA, Franka, Apptronik Apollo) are documented for the older Gemini Robotics / Robotics-ER 1.6/2 line (eth-memory/sol/reports/2026-08-08...md:69, 2026-08-13...md:74), a different model generation. Tracker appears to have carried an older/web-sourced robotics-benchmark description onto the new Argon name without a house source backing it. |
| NEURA "Neuron AI" — "launched Oct 14; 11-min laundry demo at GTC Berlin Oct 20-22"; "5 embodiments, multisensory" | CONTRADICTED | The string "Neuron AI" does not occur anywhere in the repo. House's own latest source, Logs/model-watch.md:8479 (Oct 15 17:00 scan): "NEURA demo expected but not yet confirmed in the official programme. No pre-event leaks on new product announcements." That directly contradicts a tracker claiming the demo already "launched Oct 14." NEURA Robotics itself is real and heavily sourced (Series C up to $1.4B, NEURA Gyms — eth-memory/sol/reports/2026-06-21, 2026-07-27/28, 2026-08-11), but "Neuron AI" as a named product/launch is not in any house file. |
| Nature Will (Natura-1, 50B MoE, 1M hrs first-person video, 100ms loop) | UNSUPPORTED | Zero hits for "Nature Will," "Natura-1," or variants in any house file (Sol reports, model-watch, ai-watch). No house source at all — not even a weak/old one. |
| Sharpa ($672M raise, hands) | CONTRADICTED | House files mention "Sharpa" repeatedly but only as a component supplier — "Sharpa Wave tactile hands" / "Sharpa five-fingered hands" inside NVIDIA's Isaac GR00T Reference Humanoid Robot (eth-memory/sol/reports/2026-09-19...md:37, 2026-09-20...md:29, 2026-09-23...md:33). No house file ties any dollar figure, let alone $672M, to Sharpa as a company raise. The number does not appear anywhere in the corpus in that context. |
| EU AI Office — first formal requests to OpenAI + Anthropic (Oct 14) | OK | Research/ai-watch/2026-10-15.md:19-20 and the English twin en/2026-10-15.md:19-20: Commission VP Henna Virkkunen announced at the Dublin AI summit (14.10) that the AI Office sent formal information requests to OpenAI and Anthropic — first use of AI Act powers on frontier labs, over summer incidents where test models escaped their environments. Matches the tracker's row 6 exactly, including the characterization as enforcement-grade, not just policy. |
| Unitree $904M IPO | OK | Confirmed repeatedly and consistently: eth-memory/sol/reports/2026-10-02...md:44 ("~6.1B yuan / $904M"), restated in 2026-10-08, 2026-10-13 (:44), 2026-10-16 reports — all citing AP, https://apnews.com/article/f33f…REDACTED. |
| XPENG ($900M+ raise, IRON mass prod end-2026) | OK | Confirmed: eth-memory/sol/reports/2026-10-03...md:31, 2026-10-08...md:128-130 (">$900M... post-money valuation above $6.3B"), 2026-10-13...md:40. IRON mass…REDACTED target stated consistently across all three. |
| Omdia ~18,500 units H1 2026 | OK | Confirmed across many Sol reports, e.g. 2026-10-02...md:44, 2026-10-05...md:65, 2026-10-12...md:21,50 — "Omdia estimated Chinese humanoid makers shipped about 18,500 units globally in H1 2026," cited via AP. (Note: one earlier Sol report, 2026-10-06...md:50, cites a different Omdia-sourced figure — AGIBOT and Unitree each shipping >5,000 units in 2025 — a different metric/period, not a contradiction of the H1-2026 18.5k figure, but worth knowing both numbers exist in the house record.) |
Spot-check of adjacent entity-table rows
| Claim | Verdict | Evidence |
|---|---|---|
| NVDA / Figure 100k-GPU deal on Vera Rubin (Sep) | OK | eth-memory/sol/reports/2026-09-12...md:12, 2026-09-13...md:42 — Figure + Nscale, up to 100,000 NVIDIA Vera Rubin GPUs, $3.5B initial / intent >$6B, Sept 3 2026. |
| GOOGL / "Atlas deployment committed to DeepMind" | OK, simplified | eth-memory/sol/reports/2026-10-13...md:37 — Boston Dynamics: "2026 deployments were already committed to Hyundai's RMAC and Google DeepMind." Tracker names only DeepMind; the house source splits the commitment between Hyundai RMAC and DeepMind. Not wrong, just incomplete. |
| Meta / ARI acquisition | OK | eth-memory/sol/reports/2026-10-11...md:177,241 — Meta acquired ARI for humanoid robotics AI ambitions. |
| Amazon / Fauna acquisition | OK | eth-memory/sol/reports/2026-06-07...md:65,74,132 — "Amazon/Fauna acquisition is reported by The Robot Report," consumer robotics signal. Real but from an older (June) report; tracker doesn't date it, which is fine since it's a standing position, not a dated event. |
| Figure AI — 350+ delivered, 1 robot/hour cadence, BMW + Catalyst | OK | eth-memory/sol/reports/2026-10-03, 2026-10-06, 2026-10-09 all confirm BotQ delivered 350+ Figure 03 units and raised cadence from 1/day to 1/hour. |
| Agility Robotics — S-4 effective Oct 13, vote Nov 2, $620M+, Digit 5 $300M+ orders, Toyota MFG Canada, NVIDIA Halos | OK | Verbatim match to Logs/model-watch.md:8475-8480 (Oct 15 17:00 PM scan), SEC Form 425 Oct 13. |
| Bedrock Robotics — "seeking $600M @ $3.5B pre" | CONTRADICTED | Only house figures for Bedrock: eth-memory/sol/reports/2026-07-03...md:85 — $270M Series B at $1.75B valuation. No house file anywhere mentions $600M or a $3.5B pre-money figure for Bedrock. The tracker's numbers don't match any source in the corpus. |
| Mind Robotics — "Doug Field (ex-Ford) as president; Rivian factory tests; $1B+ raised" | CONTRADICTED | House figures (eth-memory/sol/reports/2026-09-13...md:54, repeated 2026-09-23): $500M Series A + $115M seed = $615M total, ~$2B valuation, Rivian spin-out. No mention of "Doug Field" anywhere in the corpus, and "$1B+" overstates the documented $615M. The "Rivian factory tests" detail is also not stated — only that Mind Robotics is a Rivian spin-out. |
| Hesai (CN) — "Kosmo platform Oct 14 (Reuters, via Sol)" | UNSUPPORTED | Zero hits for "Hesai" or "Kosmo" anywhere in eth-memory/sol/reports/. The tracker's own changelog already flagged the HSAI ticker as unverified; this check finds the underlying Kosmo claim itself has no house source at all, not just an unverified ticker. |
| Science Corp — "PRIMA retinal implant, first Stargardt patient (May, disclosed Oct 13)" | UNSUPPORTED / likely wrong disease | eth-memory/sol/reports/2026-09-18...md:96 documents PRIMA's European commercial launch, July 22 2026, for geographic atrophy from age-related macular degeneration (AMD) — not Stargardt disease. No house file mentions "Stargardt" or a May first-patient date disclosed Oct 13. This looks like a conflation (PRIMA has also been tested for Stargardt in other trials, but not in anything in this house's record) rather than confirmed fact. |
Verdict counts
- OK: 11 (EU AI Office · Unitree $904M · XPENG · Omdia 18.5k · NVDA/Figure Vera Rubin · GOOGL/Atlas-DeepMind · Meta/ARI · Amazon/Fauna · Figure 350+/1-robot-hour · Agility S-4/vote/Digit5/Toyota/Halos)
- UNSUPPORTED (no house source found): 4 (Google Argon's specific robotics benchmarks/bodies · Nature Will/Natura-1 · Hesai Kosmo · Science Corp "Stargardt" detail)
- CONTRADICTED (a house file says otherwise): 4 (NEURA "Neuron AI" Oct 14 launch/demo · Sharpa $672M raise · Bedrock Robotics $600M@$3.5B · Mind Robotics "Doug Field"/$1B+)
Three-line verdict
The tracker's hard financial/market numbers for publicly tracked names (Unitree, XPENG, Omdia, Agility, Figure, NVIDIA/Vera Rubin) hold up and match the Sol reports and model-watch cleanly. Where it breaks down is private-company specifics and anything describing a frontier model's robotics capability in benchmark terms (Argon's ERQA/sim-lift numbers, NEURA's "Neuron AI," Sharpa's and Bedrock's and Mind Robotics' funding figures, Hesai, Science Corp's disease target) — there, several rows carry numbers or names with no house source, and at least four are directly contradicted by what the house's own files actually say. Recommend Code-Eth strike or re-source those eight rows before the tracker is treated as ground truth for portfolio decisions.
Web check 2026-10-16
Per the house rule, "no house source" meant UNSUPPORTED/CONTRADICTED above — not "false." Checked every one of those rows against the live web (primary sources: company blogs/press releases, Reuters/AP-fed wire pickups, Axios Pro, TechCrunch). All eight check out as real, current events; the tracker's author evidently worked straight from the web rather than from house files, which is exactly the gap this file exists to name, not a factual error in the tracker itself.
| Claim | Verdict | Primary source |
|---|---|---|
| Google Argon — ERQA 87%, Video-to-Sim-to-Real, 56%→75% real-world success, bodies Spot/Apollo/ALOHA 2/SO-101, GA Oct 13 | CONFIRMED | Google Developers Blog, "Gemini 4 Argon for Robotics" (Oct 13, 2026) — https://developers.googleblog.com/gemi…REDACTED/ . Exact match: ERQA 87% ("ahead of every frontier model"); Video-to-Sim-to-Real pipeline (digital-twin reconstruction + visual-reasoning feedback loop); 56%→75% success, 35s→24s task time; bodies named: Boston Dynamics Spot, Apptronik Apollo 2, ALOHA 2, SO-101 (plus Franka FR3, not in tracker). |
| NEURA Robotics "Neuron AI" — 5 embodiments, multisensory, launched Oct 14; 11-min laundry demo at GTC Berlin Oct 20–22 | CONFIRMED | NEURA Robotics press release, "NEURA Robotics Introduces Neuron AI" (Oct 14, 2026) — https://neura-robotics.com/neur…REDACTED/ . Launched Oct 14, 2026; covers five NEURA robot embodiments (4NE1 Gen 3, Gen 4, wheeled 4NE1, two bimanual platforms); adds touch/sound/memory on top of vision-language-action. GTC Berlin demo (4NE1, 11-minute laundry sequence) is scheduled Oct 20–22 at Station-Berlin Hall 3 Booth #3031 — a future event as of today (10-16), not one that already happened. House's Logs/model-watch.md:8479 (Oct 15 scan, "not yet confirmed in official programme") was simply stale relative to NEURA's own Oct 14 announcement. |
| Nature Will "Natura-1" — 50B MoE, 1M hrs first-person video, 100ms loop | CONFIRMED | Nature Will official blog, "The Will to Act — Natura-1" — https://www.naturewill.com/blog/natura-1 ; also PR Newswire, Oct 2026 — https://www.prnewswire.com/news-releases/phys…REDACTED.html . Exact match: 50B-parameter MoE transformer, >1,000,000 hours egocentric human video, 100ms control loop, drives 20-DoF dexterous hands. San Francisco-based physical-AI lab, announced ~Oct 12–13, 2026. |
| Sharpa — $672M raise | CONFIRMED (figure within FX rounding) | DealStreetAsia, "Hesai founders' Sharpa bags $670m, unveils fully automated robot restaurant" — https://www.dealstreetasia.com/stories/ai-r…REDACTED . >4.5B yuan, reported across outlets as $668M–$672.4M depending on FX snapshot (tracker's $672M sits inside that range), at a 22B-yuan (~$3.3B) post-money valuation. Investors: Alibaba, Tencent, Meituan, JD.com, Transsion, HongShan, Qiming, Luminous. Note: this round was announced ~Aug 30, 2026 (tied to the Dairy Queen robot-ice-cream-store launch), not a new Oct event — the dollar figure is right, the date in any "Oct 14" framing would not be. |
| Bedrock Robotics — "$600M @ $3.5B pre" (house file only had Feb's $270M @ $1.75B) | CONFIRMED, reported/in-progress | Axios Pro, "Bedrock Robotics raising $600M for autonomous construction robots" (Oct 13, 2026) — https://www.axios.com/pro/climate-deals/2026/10/13/bedr…REDACTED ; also thisweekinvc.substack.com, same figures. Bedrock is reportedly raising a new Series C at a $3.5B pre-money valuation — 2x the Feb Series B — on top of (not instead of) the $270M/$1.75B round the house file has. Both figures are real; they're sequential rounds, not a contradiction. Flagged as "reportedly raising" by Axios, i.e. not yet confirmed closed as of Oct 13. |
| Mind Robotics — "$1B+ raised" / Doug Field as president (house file only had $615M, no Doug Field) | CONFIRMED | Electrek, "Rivian's Mind Robotics snags Doug Field, ex-Tesla and Ford EV chief" (Oct 13, 2026) — https://electrek.co/2026/10/13/rivi…REDACTED/ . Doug Field (ran Tesla's Model 3 program, later Ford's EV unit) is now president, running product/engineering/ops alongside RJ Scaringe. Total funding: $115M seed + $500M Series A (March 2026, the house file's $615M figure) + a further $400M Series B led by Kleiner Perkins (disclosed since, valuing Mind at $3.4B, up from $2B two months prior) — crossing $1B in under a year. House figure was simply pre-dated by a round the house hadn't logged yet; "Rivian factory tests" is consistent with reporting that Rivian's Normal, IL plant serves as Mind's live training environment, though no outlet uses that exact phrase. |
| (bonus, flagged but not in the named list) Hesai "Kosmo" platform, Oct 14 | CONFIRMED | unite.ai, engineering.com, technology.org, ksl.com (AP-fed), all dated Oct 14–15, 2026. Kosmo = a real product launch: Kosmo One spatial camera + Kosmo World cloud platform, a 3D world-data engine for robot training, launched by Hesai (whose three co-founders also founded Sharpa) on Oct 14, 2026. Direct Reuters byline not independently located, but the event, name, and date are solid across independent outlets — the house file's "zero hits" reflects a web-only announcement, not a fabrication. |
| (bonus) Science Corp PRIMA — first Stargardt patient (May, disclosed Oct 13) | CONFIRMED | TechCrunch, "Science Corporation implants its vision-restoring chip in its first Stargardt disease patient" (Oct 13, 2026) — https://techcrunch.com/2026/10/13/scie…REDACTED/ . Confirms exactly: implanted May 2026 (Sydney Eye Hospital trial), disclosed via TechCrunch Oct 13, 2026. This is a separate PRIMA trial from the AMD/geographic-atrophy European commercial launch the house file had on record (eth-memory/sol/reports/2026-09-18...md:96) — both are real, they're different patient populations, not a contradiction. |
Two-line verdict: All eight web-sourced rows the house check could not back are real, current, and essentially accurate — they came from the web rather than from house files, which is the gap the verify file was built to name, not an error in the tracker. The two numeric "CONTRADICTED" calls (Bedrock, Mind Robotics) only contradicted stale house figures that predate a second funding round each company has since announced or closed; recommend updating the house record with these rounds rather than striking the tracker rows.
Source in the house: Research/physical-ai-tracker-verify-2026-10-16.md& Ethan