Overview of #292 Brett Adcock — Shawn Ryan Meets a Humanoid Robot
This episode features Brett Adcock (founder/CEO of Figure, co‑founder of Archer Aviation, founder of Cover, and founder of Hark), interviewed by Shawn Ryan. The conversation covers Brett’s entrepreneur origin story, the technical and commercial progress of Figure’s humanoid robots (Figure 1–3 and the Helix neural stack), Archer’s eVTOL work, a terahertz-based weapons‑detection venture (Cover) to protect K–12 schools, and a new AI lab/hardware initiative (Hark). The episode mixes engineering detail, business strategy, deployment plans, safety and ethics, and advice for founders.
Major topics covered
- Brett Adcock’s background: small‑town Midwest roots → University of Florida → web startups → Vetteri (AI recruiting marketplace; acquired) → Archer (eVTOL) → Figure (humanoid robots) → Cover (school weapon detection) → Hark (AI + hardware).
- Figure (humanoid robotics): hardware design, neural‑net-first software (Helix), demonstrations, commercial pilots (BMW, logistics partner, Brookfield), manufacturing cadence and scale targets.
- Archer Aviation: eVTOL design choices, certification challenges (FAA), and the vision for urban air mobility.
- Cover: terahertz/millimeter-wave weapon detection spun out from JPL tech, goal to reduce school shootings via standoff concealed weapon detection, cost reduction from research prototype to production chips.
- Hark: new AI lab building multimodal models and next‑generation devices (post‑phone computing) focused on human‑centric AI and personal automation.
- Safety, ethics, and societal impacts: physical safety around humanoids, data & permissioning, military use choices, and the balance of risk vs. upside from automation.
Key takeaways and highlights
- Figures of scale and milestones
- Figure 3 robot: ~130–135 lb, ~5'6", ~40 degrees of freedom (≈40 actuated joints); palm cameras + tactile fingertip sensors; picks ~40 lb boxes; runs 4–5 hours per charge; ~1 hour to charge (inductive charging via feet).
- Manufacturing: current line can produce a robot every ~90 minutes; facility peak stated capacity around 40,000–50,000/yr; ambition to reach ~1 million units/yr over the decade.
- Funding: Brett has raised very large rounds historically (mentions ~$700M Series B led by OpenAI and Microsoft, and raising in the billions overall; Figure later referenced ~$2B raised).
- Software strategy: moved from hybrid code + heuristics to predominantly neural‑net controllers (Helix / Helix 2). Removing ~100k lines of bespoke code and replacing with neural policies improved generalization and reduced brittle failures.
- Real‑world deployments: robots have operated months in commercial settings (BMW body shop tests, logistics pilots) and run 24/7 shifts with robot→robot handoffs for charging and fault handling.
- Commercial → Consumer rollout logic: deploy to lower‑variability, higher‑ARPU commercial settings first (manufacturing, logistics, facility greeters). Home deployment is harder due to variability, safety/trust, and pricing constraints; consumer price target discussed around the order of $500/month subscription-style rather than an upfront cost.
- Safety & trust: multiple layers — intrinsic hardware safety (soft skins, safe actuators), perception and semantic safety (avoid knocking over hot liquids/candles), authentication/permissioning for actions (voice + facial recognition/fingerprint for sensitive commands), and extensive testing before household autonomy near children.
- Cover (weapons detection): terahertz imaging can detect concealed weapons in backpacks/waistbands at standoff ranges; goal to cut costs dramatically by custom chips (research prototypes had expensive hardware; they designed chips to get BOM down); target beta testing in schools within the year.
- Hark (AI lab + devices): building multimodal models and new hardware that act as always‑on, personalized assistants (“Jarvis”‑like vision), with the intent to replace old phone/PC interaction models.
Technical & product detail (Figure)
- Hardware:
- Electric actuators designed for humanoid dynamics (motors, gearbox, sensors, motor controllers, thermal management).
- Hands: 5th‑generation hands, cameras in palms, tactile sensors on fingertips; sensitivity down to small forces; forthcoming hand designs approach human joint counts.
- Torso houses batteries, GPUs, main compute and power distribution.
- Inductive charging pads for feet; robots swap in/out autonomously for charging during continuous shifts.
- Software:
- Helix neural‑net stack for perception, planning, and control. Recent shift to neural controllers for walking and manipulation (controllers run ~200 Hz; low‑level motor loops run up to 5–6 kHz).
- Training using large, diverse datasets; ability to load new “apps” (weights) to repurpose robots for logistics, laundry, etc.
- Operations:
- Robots communicate via onboard 5G (eSIM, T‑Mobile) and coordinate robot→robot for docking/coverage.
- Monitoring and remote/edge autonomy; robots can limp to triage when hardware faults occur.
- Demonstrations:
- Keurig K‑cup demo (neural net perception + manipulation).
- BMW pilot: Figure units performed repetitive body‑shop tasks 10‑hr shifts daily for months.
- Office greeters and logistics deployments running continuous service.
Archer Aviation — eVTOL recap
- Approach: electric vertical takeoff & landing (eVTOL) aircraft with distributed electric propulsion to increase redundancy, reduce acoustic signature, and lower cost.
- Challenges: electric energy density (battery vs kerosene), design tradeoffs between rotor disk area and motor redundancy, FAA type certification with extremely stringent safety goals (target catastrophic rate ~1×10^‑9 hours).
- Status: flown weekly in California; public company via SPAC; certification and regulatory approval remain long‑pole items.
Cover — weapons detection (summary)
- Tech: terahertz / millimeter‑wave radar imaging (high‑frequency RF imaging) producing point clouds/2D images that reveal concealed objects through clothing and backpacks.
- Origin: tech originated at NASA JPL; Adcock spun a team out, co‑developed custom chips to reduce cost dramatically (research hardware expensive; goal chips cost much lower).
- Goals & deployment: make a standoff, noninvasive, high‑frame‑rate detector that can scan school entrances without stopping kids; reduce false positives, avoid frightening students, and make the system affordable for wide K–12 rollout. Beta testing with real schools intended.
- Broader use cases: stadiums, airports, hospitals, malls, other public venues.
Hark — human‑centric AI + new hardware
- Vision: build the next computing platform for AI — multimodal models + always‑on device(s) that know and act for you (personal operating system, far beyond current chatbots).
- Approach: self‑funded lab, hiring top AI and design talent (including lead iPhone designer), building models and new hardware form factors to deliver a “Jarvis”‑like memory, tool use, and action in the digital & physical stack.
- Positioning: aims to be complementary but distinct from existing large‑model labs; focus on privacy, persistent memory, tool use, and new interfaces.
Safety, security, privacy, and ethics
- Safety engineering is central: intrinsic mechanical safety (soft coverings, safe actuators), redundant motors, and conservative operating envelopes.
- Semantic safety: perception models to avoid dangerous actions (not knocking over candles, avoiding boiling water spillage), authorization systems for actions that can spend money or affect children.
- Authentication: multi‑factor expectation — voice recognition is not enough for sensitive acts (use facial recognition, biometrics, permissions).
- Military: Figure is not pursuing military applications at present; Brett cited complexity and divergence of focus as reasons.
- Risks acknowledged: malfunctions, hacks, misuse, and societal disruption — but Brett is optimistic that the productivity benefits and reductions in drudgery outweigh risks if steered rightly.
Notable quotes
- “We have more positions the body could be in than atoms in the universe” — (on combinatorial complexity of humanoid pose space).
- “The hardest thing in the stack is can you put a robot into your home today and do the five hours of work you need without ever seeing your home before.” — (on consumer deployment challenge).
- “If you solve general‑purpose robotics, you can ship billions of robots.” — (on the scale and market opportunity).
- “We’re building synthetic humans at scale.” — (on the combined effect of embodied and digital AI).
Actionable points / recommendations (for different audiences)
- For enterprise customers (logistics, manufacturing, real estate): consider pilots now — humanoid automation can solve compliant, variable manipulation tasks and run continuous shifts; engage with vendors to define KPIs and tolerances.
- For school administrators / policymakers: monitor Cover’s beta work for potential noninvasive standoff detection solutions; evaluate tradeoffs (privacy, false positives, cost).
- For product builders & founders:
- Pick ambitious, high‑upside problems (hard problems attract talent and capital).
- Start building early; iterate quickly; learn from real deployments.
- Expect deep technical challenges and long time horizons for robot + hardware businesses; plan financing and team accordingly.
- For the public: understand deployment will be incremental — commercial/industrial first, then homes — and that safety, permissioning, and regulation will shape timing.
Quick reference (figures & claims mentioned)
- Figure 3: ~130–135 lb, 5'6", ~40 DOF.
- Battery runtime: ~4–5 hours; charge ~1 hour.
- Manufacture cadence: ~1 robot per 90 minutes on running line; facility peak stated 40k–50k/yr; goal ~1M/yr in the decade.
- Neural stack: Helix; shift to almost entirely neural controllers (removed ~100k lines of code).
- Funding/financing notes: Vetteri sale ~ $110M; Series B ~ $700M led by OpenAI & Microsoft; Brett references raising billions overall; Hark self‑funded with ~$100M (per intro).
- Archer safety target: catastrophic failure rate ≈ 1×10^‑9 hours.
Final assessment
This episode is a deep, wide‑ranging conversation blending engineering depth and founder strategy. Brett Adcock lays out tangible progress (working robots in continuous operation, commercial pilots, manufacturing lines) and near‑term goals (commercial rollouts, Cover beta tests, and Hark productization). He is optimistic about the transformative potential while acknowledging safety, regulatory, and societal hurdles. For listeners wanting to understand where humanoid robotics and related AI systems actually stand (not the sci‑fi headlines), this is a valuable, grounded update from a founder operating across hardware, AI, and regulated industries.