Duke of AI
Duke's Tech Knowledge Database Living Handbook · v13 · Compiled 15 September 2026

Space, Robotics & Emerging Hardware

Core Principles

  • A recurring framing across source material: AI, robotics, and reusable space technology are described as mutually reinforcing rather than independent trends.
  • Progress in humanoid robotics is currently driven more by incremental improvements across many components than by a single breakthrough.
  • Reliance on cloud-hosted frontier AI carries business risk, increasingly cited as a reason to evaluate local/on-device AI hardware as a complement rather than a replacement (DK-6).
  • Beyond raw hardware capability, a parallel and arguably harder challenge is interaction design — making robots "legible" and responsive to human social signals (DK-22).
  • The pursuit of human-like robot appearance is itself contested: a more human-like form raises expectations of human-level competence, which can backfire when the robot falls short (DK-22).
  • "Physical AI" is emerging as a distinct discipline from language/image AI, with the primary bottleneck described as a "robot data gap" (DK-26).
  • Commercial humanoid companion robots marketed to combat loneliness are now a real, priced product category, raising the same disclosure/dependency concerns documented elsewhere for AI content and companionship (DK-25).
  • Fleet-scale collective learning is a stated competitive strategy for humanoid robotics (DK-34).
  • Humanoid robot pricing is bifurcating sharply, from ultra-low-cost developer platforms to flagship research/industrial platforms from the same manufacturer (DK-35).
  • Livestreamed, unedited long-duration demonstrations are emerging as a credibility-building format for humanoid robotics companies, though they invite the same skepticism about genuine autonomy that has surrounded earlier robot demonstrations (DK-36).
  • Power efficiency, not just raw compute, is emerging as the binding constraint for humanoid robot AI hardware (DK-53).
  • Speculative large-scale infrastructure projects are increasingly discussed alongside near-term product announcements, even when the source itself flags them as long-term ambitions rather than committed plans (DK-50, DK-53).
  • NEW (v10) — A distinct "intelligence layer, not hardware" strategy is emerging in humanoid robotics, mirroring this Handbook's existing "two AI races" framework for Apple: rather than building its own robot body, Google DeepMind is positioning its Gemini Robotics models as a portable brain that runs across other manufacturers' hardware — the same checkpoint demonstrated on multiple robot bodies from different makers, with an on-device version that adapts to a new robot in hours on fewer than 200 demonstrations (DK-78). This is a meaningfully different competitive bet than Tesla's or Figure's vertically-integrated hardware-plus-software approach documented elsewhere in this topic.
  • NEW (v10) — Government trade policy has become an explicit, named lever in the humanoid robotics competition, not just a background factor: a national-security-grounded US import ban targeting foreign-made humanoid robots, robot dogs, and related devices is aimed squarely at China's dominant position in the consumer/humanoid robotics market — extending this Handbook's existing observation about China's manufacturing scale (DK-26) into active policy response.
  • NEW (v11) — In humanoid robotics the ceiling argument stays fixed while the floor keeps being restated, and this Handbook can now show it rather than assert it. Two interviews with the same Figure AI founder ten months apart (June 2025, DK-90; April 2026, DK-88) keep the addressable-market claim identical — a little under half of world GDP is paid as human labour — while the near-term shipping figure falls from 100,000 robots within four years to "thousands" this year, the flagship BMW deployment moves into the past tense, and the stated bottleneck moves from manufacturing to intelligence. Treat forward-looking humanoid volume targets accordingly.
  • NEW (v11) — The bottleneck in humanoid robotics is now stated by a leading vendor as intelligence rather than manufacturing — "this is not a manufacturing problem, this is an intelligence problem" — with commercial demand described as far exceeding what can be reliably shipped. This inverts the earlier framing in which production scale was the limiting factor, and is consistent with this Handbook's existing "robot data gap" principle (DK-26, DK-88).
  • NEW (v12) — Apple's release structure changed, not just its hardware: for the first time it launched Pro models with no base iPhone beside them, with the standard model said to be coming the following year, while raising the Pro Max another $100 for a device externally unchanged and topping the range at $3,199 (DK-94). Read together, that is a company moving its volume line and its halo line further apart.
  • NEW (v13) — The test for a humanoid robot is manipulation, not locomotion. A robot that runs, backflips or dances in a controlled environment tells you nothing about whether it can pick up an egg without breaking it, open a fridge or control a stove — and cooking an egg is harder for a robot than a rehearsed backflip (DK-108). Apply this to every humanoid demonstration video: watch the hands, discount the athletics.

Key Facts & Examples

  • 1X Robotics unveiled a robotic hand with 25 degrees of freedom, tendon-driven, enabling waterproof operation (DK-2). NEW (v10) — Proception, founded by a former Tesla Optimus technical lead who settled a trade-secret suit with Tesla, raised an $11 million seed for its own 22-degree-of-freedom robot hand plus a sensor-packed data-capture glove, reinforcing this Handbook's standing observation that hands remain one of the hardest parts of humanoid robotics to get right (DK-70).
  • China's Long March program achieved a successful reusable-booster landing; SpaceX was described as still maintaining a significant operational lead (DK-2).
  • Local AI hardware named as increasingly viable: NVIDIA DGX Spark, Framework Desktop, Apple Mac Studio (DK-6).
  • Factories and warehouses are consistently identified as the realistic first deployment market for humanoid robots (DK-26). China is described as establishing a leadership position in humanoid robotics through strong government support and integrated manufacturing (140+ companies, ~$140 billion pledged 2025 investment) (DK-26).
  • Tesla's reported Optimus strategy: a fleet of 10,000–30,000 robots learning from real-world experience via "Optimus Academy"; Gen 3 hands reported at 22 degrees of freedom (DK-34, corroborated DK-51).
  • Unitree's humanoid lineup spans R1 (~$4,900) through flagship H1 (~$90,000) (DK-35).
  • Figure AI livestreamed three Figure 03 humanoids performing an unedited 24+ hour warehouse shift, sorting 30,000+ packages (DK-36).
  • Tesla's AI5 chip completed tapeout at Samsung's Taylor, Texas plant, reportedly optimized to draw ~250W (down from 500–800W estimates) to extend Optimus's operating time on its 2.3 kWh battery (DK-53).
  • SpaceX's Starlink business has scaled to over 10 million subscribers across 160+ countries, reportedly generating $11.4 billion of SpaceX's $18.7 billion total 2025 revenue (DK-52).
  • SpaceX's Starship Flight 13 (Starship V3) is designed to prioritize reliability over altitude, for the first time carrying 20 operational Starlink V3 satellites rather than dummy payloads (DK-53).
  • NEW (v10) — SpaceX began trading on 12 June 2026, raising $75 billion at a $1.75 trillion valuation — described in the source as the largest stock offering ever. In the two months prior, SpaceX had signed compute deals with Anthropic and Google worth roughly $26 billion a year combined (inherited from the xAI merger and its Colossus data centre in Memphis); notably, Google — one of the largest owners of AI compute in the world — is reported to still be renting capacity from SpaceX to meet Gemini demand. The underlying thesis is that the next major compute platform will be in orbit rather than on Earth, where solar power is constant and cooling is free — which depends on three still-unproven pieces at once: fully reusable Starship launches, the proposed Terafab chip foundry (already flagged as a long-term ambition rather than a committed plan — DK-50), and a satellite factory intended to produce roughly 556 AI satellites a month that does not yet exist. Two independent analyst valuations landed well below the bankers' figure: Morningstar at ~$780 billion, NYU's Aswath Damodaran at ~$1.2 trillion (DK-66). AMD's separately reported plan to sell Anthropic tens of billions in AI servers, alongside Anthropic's existing rental of SpaceX's Colossus 1 facility, adds further texture to how tightly AI compute demand and SpaceX's business case are now linked (DK-76).
  • NEW (v10) — Google DeepMind released Gemini Robotics 2, coordinating a robot from feet to fingertips off a single instruction — balancing, walking, bending, and handling delicate objects from one model. Reported highlights: 22 degrees of freedom in the hands (tying trash-bag knots, sealing ziplock bags, unscrewing lightbulbs at a claimed 92% success rate on Apptronik's Apollo 2), the same checkpoint running on different robot bodies including Franka's bi-arm platform, and an on-device version that reportedly adapts to a new robot in hours on fewer than 200 demonstrations. Dexterity on complex tasks is still reported at only 30–45%, and the robots remain slow (DK-78).
  • NEW (v10) — The Trump administration is banning imports of new foreign-made humanoid robots, robot dogs, robot vacuums, and power inverters on national-security grounds; the FCC's "advanced robotic devices" definition extends to quadrupeds and many ground-traveling robots with wireless connectivity and sensors, potentially sweeping in future robot lawnmowers, sidewalk delivery, and warehouse robots. Existing already-approved devices are unaffected. China's Foreign Ministry reportedly said it would use all measures necessary to protect its businesses (DK-78).
  • NEW (v10) — Zoox is reported to be ready to start charging for robotaxi rides in Las Vegas, following a federal exemption for its steering-wheel-free, pedal-free vehicles — a notable regulatory precedent for fully custom-designed (not retrofitted) autonomous vehicles (DK-79).
  • NEW (v11) — Figure AI, June 2025 baseline (DK-90): BotQ described as having installed capacity of 12,000 robots per year per line, with a stated target of 100,000 robots shipped within four years, justified by reclassifying the problem as consumer-electronics rather than car manufacturing. BMW named as the deployment, doing body-shop work moving sheet metal; a second logistics customer signed but unnamed, which the interviewer said Bloomberg had reported as UPS. Figure 3 claimed to be "90% cheaper than Figure 2, like 93%". No teleoperation in deployed work, used only for data collection and testing.
  • NEW (v11) — Figure AI, April 2026 (DK-88): near-term output now given as "thousands of robots" this year with record production in March and an intent to triple it by May; a million units a year is described as a later ambition. The BMW deployment is described in the past tense — a small batch ran daily for six months, prompting a refactor of the whole commercialisation approach that produced Helix 2, their second-generation model. Reliability progression given as Figure 1 faulting after about an hour, Figure 2 roughly once a day, Figure 3 running all day with faults seen weekly; he notes the honest corollary that absolute fault counts rise as the fleet grows. His stated bar is a robot doing seven to ten hours of useful home work with no human intervention, daily — "nobody's ever shown that".
  • NEW (v11) — The collective-learning thesis this Handbook already records as a stated strategy (DK-34) is put in its strongest form by Figure: because robots share a neural network and improve together, whoever fields the most useful robots gets both the cheapest and the smartest, which he argues could make humanoid robotics "a winner take all market" — unusual for advanced hardware. Recorded as a competitive thesis from an interested party, not a finding (DK-90).
  • NEW (v11) — Figure's stated differentiator is vertical integration to an unusual depth: motors, rotors, stators, sensors, structure, kinematics, joints, batteries and packs all designed in-house, on the reasoning that otherwise "you're left at the mercy of some vendor". This is the opposite bet to Google DeepMind's portable-brain approach recorded above (DK-78, DK-88).
  • NEW (v11) — On the OpenAI relationship, worth recording because the same event is told two ways to two audiences. In June 2025, diplomatically: "we were just better at doing it alone", plus a reputational argument that partnering led people to assume OpenAI had done the work (DK-90). In April 2026, bluntly: OpenAI led the Series B and brought in Microsoft, his team ended up "running circles around" them, "so I fired them", and of the original strategic rationale, "it turned out I was kind of wrong on that" (DK-88). Helix is confirmed to use an open-source vision-language backbone for semantic grounding, with data collection, training and inference in-house.
  • NEW (v12) — Apple's iPhone Duo is its first foldable: a passport-style fold, only slightly thicker than the 18 Pro Max when closed and the thinnest iPhone ever when open, with Apple's first under-display camera and Apple Pencil support. The concession is authentication — no Face ID, only Touch ID in the power button, on a phone starting at $1,999 (DK-94). The camera change worth noting across the range is a physically adjustable aperture on the main wide lens, described as the biggest iPhone camera change in four years; the 48MP sensor is otherwise essentially unchanged. All performance figures are Apple's own — the presenter did not attend the event or handle the devices.
  • NEW (v12) — Tesla's Cybercab event put the stated intention at about $30,000 a unit, so a buyer can run several as revenue-earning robotaxis, with parts of cities expected to exclude human drivers on efficiency grounds (DK-91). A stated intention, not a shipped price.
  • NEW (v13) — Optimus Gen 3's hands are reported at 22 degrees of freedom against 11 on the previous generation, with close to three times the internal sensors and possible water resistance. The design choice worth recording is that most actuators move up into the forearm, with tendon-like cables through the wrist driving the fingers — the arrangement human muscles and tendons use — keeping the hand light and precise while the force comes from behind it. Claimed to lift around 40 lb / 18 kg while still handling an egg (DK-108).
  • NEW (v13) — Optimus pricing, and it corrects the figure most coverage repeats: Wall Street forecasts and robotics experts are reported putting early units potentially at $70,000, with $50,000 called a reasonable possibility, falling later if production scales (DK-108). The familiar $20,000-$30,000 is a target, not a price. Separately, one source puts a fully capable domestic robot bottoming out around $20,000 with five-finger hands costing about $10,000 each (DK-111).
  • NEW (v13) — The AI5 computer is rumoured at roughly five times the memory bandwidth of AI4, and the argument made from it is more interesting than the number: the bottleneck in humanoids is not processor power but moving data from memory to processor fast enough, so the gain removes the "stop and think" pauses rather than making the robot cleverer (DK-108). The proposed architecture is two-layer — perception, balance and motion local to the robot, with Grok as a higher-level language and coordination layer.
  • NEW (v13) — Cybercab figures from 2026 EPA certification, which makes them firmer than keynote numbers: about 293 miles real-world range, a 48 kWh pack of 4680 cells, 165 Wh per mile, drag coefficient 0.20, 219 hp, 3,113 lb, and 20.2 cubic feet of cargo with no front trunk. Two genuine firsts for Tesla: a rare-earth-free motor — rare earths are said to be roughly 25% of an electric motor's cost — and fully integrated brake-by-wire with no hydraulic link from pedal to caliper (DK-108).
  • NEW (v13) — Starship status as at September 2026, from Musk: flight 14 is the last before attempting to catch the ship, flight 15 is the attempt, with reflight of both ship and booster either late this year or more likely early next. He puts the odds of catching it first time at 50-60%, noting the previous flight's simulated landing would have been caught had a tower been there, and says full rapid reusability in 2027 is "extremely likely" (DK-112).
  • NEW (v13) — Why full reusability is the threshold rather than a milestone, in his framing: the Shuttle was partly reusable but so hard to reuse that it cost more per flight than an expendable rocket, and Falcon 9 still loses its upper stage every time — "about the cost of a medium-sized jet". Starship returns both stages to the pad and is designed for aircraft-like turnaround (DK-112).
  • NEW (v13) — The case for data centres in orbit, argued from permitting and economics rather than spectacle, which makes it testable: ground real estate reprices the moment a data centre is announced — $3,000 an acre to $180,000 is the figure cited — permits are slow, and generators are quoted at three years. In orbit the real estate is free, cooling radiates to deep space, and satellites can be oriented to face the sun continuously. SpaceX owns launch, which is the part competitors must buy. AI compute satellites are said to be launching next year (DK-112).
  • NEW (v13) — Terafab's stated rationale is supply risk rather than ambition: chips from Taiwan may at some point not arrive, and separately every existing fab is running at maximum capacity, so scaling AI into robots and cars runs out of room regardless of geopolitics. "It's either build Terafab or fail to scale." Status is an R&D line at Austin with equipment on order, expecting something useful by end of next year — explicitly framed as crawl, walk, run (DK-112).
  • NEW (v13) — Robot production ramp, as claimed by one source and worth logging because it carries dates: 11,000 humanoid robots produced last year, millions within a few years, tens of millions by 2030, a billion in 10-15 years — against roughly 70 million cars and 70 million motorcycles built annually today (DK-111).

Source Articles: DK-2, DK-6, DK-22, DK-25, DK-26, DK-34, DK-35, DK-36, DK-48, DK-49, DK-50, DK-51, DK-52, DK-53, DK-66, DK-70, DK-76, DK-78, DK-79, DK-88, DK-90, DK-91, DK-94, DK-108, DK-111, DK-112

Version 13, compiled 15 September 2026. 112 source entries (DK-2 – DK-113). Claims are recorded as their sources framed them; figures described as claimed or reported are not independently verified.