Applied Manipulation and Deployed Robotics · Embodied Intelligence in 2026

6.0 Why this chapter matters more than its funding suggests

The companies in Chapters 4 and 5 have raised, collectively, well over $20B. The companies in this chapter have raised a fraction of that and are the only ones with unambiguous evidence that robots are doing paid work at scale. If the data flywheel thesis is correct — that deployment generates the data that generates the intelligence — then this is where the flywheel is actually spinning.

Read each entry for one thing: what is the data loop, and is it a manipulation loop or a logistics loop? The difference decides whether these companies are quietly accumulating the corpus everyone else is trying to buy, or merely accumulating telemetry.

6.1 Dexterity AI

Thesis: physical AI must be earned in production, not in simulation. Build task-specific superhuman manipulation for freight, then generalise upward from deployment logs.

Deployed: DexR multi-arm parcel handling and Mech, a dual-armed mobile manipulator for truck and trailer loading, launched March 2025 and manufactured with Sanmina. Customers include FedEx (autonomous trailer loading), a Fortune 100 retailer, and Sagawa Express.

Data: the clearest deployment-to-model flywheel in the sector. In March 2026 Dexterity announced Foresight, a "physics-consistent world model" the company states was trained on experience from over 100 million autonomous actions in production. It runs a 4D box-packing agent deciding in under 400 ms, spanning six applications, four robot types and five hand types.

Embodiment: anthropomorphic but not humanoid — Dexterity markets Mech as an "industrial superhumanoid": arms and torso where useful, wheels not legs.

Funding: $95M in March 2025 at a $1.65B valuation, after a $140M Series B.

Skeptic: freight loading is a narrow, high-cycle-time niche. 100 million "actions" are heavily correlated box picks, and Foresight's generality is a marketing claim not yet independently benchmarked.

6.2 Ambi Robotics

Thesis: a Berkeley spin-out from Ken Goldberg's group, betting that simulation pretraining plus continuous fleet data beats hand-engineering for parcel induction and sortation.

Deployed: AmbiSort parcel sorters, AmbiStack mixed-case palletising (January 2025), AmbiVision (March 2026), and an AI Skill Suite on AmbiOS. Anchor customer Pitney Bowes since 2021, plus OSM Worldwide; a 2026 integration with Pickle Robot links truck unloading to palletising.

Data: PRIME-1, a warehouse foundation model announced January 2025, trained on years of real production picks. Goldberg's own Science Robotics piece cites Ambi as having accumulated 22 years of real robot data in four years while sorting 100M+ packages — the single best existence proof for the deployment-flywheel thesis, and notably it comes from the field's loudest data-scarcity skeptic.

Embodiment: aggressively non-humanoid — gantries and fixed arms with suction, optimised for throughput per square foot.

Skeptic: parcel sortation is commoditising, and Ambi is smaller than Dexterity or Nimble. No verified 2025–2026 funding round was found; treat later funding claims as unverified.

6.3 Path Robotics

Thesis: welding is the highest-value labour shortage in heavy industry, and it is fundamentally a perception problem, not a motion-planning problem.

Deployed: AW-2 autonomous welding cells and Rove, a mobile welding system, sold into fabrication shops, defence and shipbuilding — including a deal with America's largest military shipbuilder.

Data: unusually rich. Path scans each part in situ, generates the weld path, and closes the loop on the resulting bead — laser scans, seam geometry and inspection outcomes on parts that are never identical, which is precisely the regime where teach-pendant programming fails. In 2025 it launched Obsidian, a foundational AI model for welding.

Funding: $100M Series D (October 2024); the company reported surpassing $100M in bookings in 2025.

Skeptic: welding data transfers poorly outside welding. The moat is vertical, not a step toward general manipulation, and defence certification cycles are slow.

6.4 Chef Robotics

Thesis: the loudest counterargument to simulation-first robotics. Food is deformable, organic and variable, so the only path is real production data at volume — and the way to get it is to charge customers for it.

Deployed: ChefOS-driven ingredient-portioning arms in high-mix food production, sold as RaaS, in more than a dozen facilities across the US, Canada and Europe since Amy's Kitchen in 2022.

Data: Chef reached 100 million meal servings in April 2026 — 1M in April 2023, 10M in January 2024, 50M in May 2025. CEO Rajat Bhageria's public argument is that RaaS deliberately trades margin for a data flywheel: start with high-volume, lower-complexity tasks, then climb.

Funding: $43.1M Series A (April 2025) led by Avataar Ventures.

Skeptic: 100 million servings is 100 million scoops of a handful of ingredient classes — impressive throughput, questionable diversity. The claim that it exceeds all other food robotics data combined is Chef's own and unaudited.

6.5 Nimble Robotics

Thesis: sell the whole fulfilment centre, not a picking arm. Vertical integration lets you collect data across storage, retrieval, pick, pack and sort rather than one station. Nimble was also the archetypal teleoperation-as-training-signal company: remote human pilots handled exceptions, and each intervention was a labelled demonstration.

Deployed: a full autonomous fulfilment system plus a Cloud Logistics Platform; CEO Simon Kalouche claims it replaces "over a dozen individual pieces of equipment" and removes as much as 70% of cost.

Funding: $106M Series C at a $1B valuation, co-led by FedEx alongside a commercial agreement.

Skeptic: the honest metric for a human-in-the-loop company is the intervention rate, and Nimble does not publish it. Owning the whole stack means capital intensity and slow deployments.

6.6 Collaborative Robotics (Cobot)

Thesis: Brad Porter (ex-Amazon Robotics VP) argues the humanoid form is a distraction. Solve the dispatch problem — knowing what work exists — and a wheeled manipulator captures most of the value.

Deployed: Proxie, a mobile cart-mover; Proxie Gen 2 (June 2026) tows 1,500 lb carts, lifts 220 lb, uses 40% fewer parts, and adds an optional dual-arm configuration. Customers include Mayo Clinic and Maersk; deployments start at $5,000/month.

Data: the standout claim is autotasking — multimodal models build a live map and infer when material needs moving without WMS integration. At Maersk, roughly 95% of cart movements happened without human task assignment. Twenty-eight Gen 1 units logged nearly 13,000 operating hours feeding fleet learning.

Funding: $30M Series A (2023), $100M Series B (April 2024). Later rounds unverified.

Skeptic: 28 units and 13,000 hours is a pilot, not a fleet, and the manipulators remain research-grade.

6.7 Diligent Robotics

Thesis: hospitals are the best-instrumented, most labour-starved indoor environment in the developed world, and a socially-aware mobile manipulator can accumulate operational hours there faster than anywhere else.

Deployed: Moxi, a one-armed mobile robot doing pharmacy, lab and supply deliveries. Moxi 2.0 (October 2025) was "built for AI." Partnership with Swisslog Healthcare; expansion into senior living.

Data: 1 million picks by February 2025 and 300,000 pharmacy deliveries by July 2025 — mostly navigation, elevator and door interaction, and constrained grasps, deep in human-populated corridors.

Embodiment: deliberately one arm and a head — social legibility over dexterity.

Outcome: in January 2026 Serve Robotics agreed to acquire Diligent. Ten years in, unit economics were never proven publicly.

6.8 RightHand Robotics

Thesis: the oldest pure-play piece-picking bet — a "model-free" grasping stack (hybrid suction and finger gripper plus learned grasp proposal) should pick millions of unseen SKUs without per-item modelling.

Deployed: RightPick cells for e-commerce, pharmaceutical, apparel and grocery fulfilment, integrated with AutoStore, AS/RS, AMRs and sorter induction, at cycle times as fast as ~3 seconds per pick. European customers include apo.com Group and Apotea.

Data: every pick produces a grasp-attempt-and-outcome record across an enormous SKU tail — arguably the cleanest large-scale grasp dataset in commercial robotics. RightHand publishes almost nothing about how it uses this for model training, which is a real transparency gap.

Funding: roughly $100M+ historically; in March 2025 Rockwell Automation took a strategic investment and became a distribution partner — a partial-exit signal.

Skeptic: the category has been "about to break out" since 2018, and RightHand has been overtaken in narrative by newer entrants.

6.9 Intrinsic (Google)

Thesis: the bottleneck in industrial robotics is not hardware or even policy quality but programming cost. A hardware-agnostic software layer that lets non-experts specify tasks unlocks the installed base.

Deployed: Flowstate, a visual robot-application platform; an "Intelligence Cell" (June 2026) aimed at eliminating manual robot coding. Intrinsic owns Open Source Robotics Corp (ROS/Gazebo stewardship), partners with NVIDIA on Isaac Manipulator, and announced a Foxconn collaboration in November 2025.

Data: breadth over depth — telemetry and task specifications across many customers' heterogeneous cells, rather than one deep vertical stream.

The decisive fact: in February 2026 Intrinsic left Alphabet's Other Bets and joined Google, with Wendy Tan White reporting to Hiroshi Lockheimer and the team working closely with DeepMind on Gemini integration. Read charitably that is consolidation of Alphabet's physical AI. Read skeptically it is absorption after five years and several strategy resets without disclosed commercial traction.

6.10 The warehouse incumbents: Symbotic and Mytra

Thesis: most warehouse value comes from case- and pallet-level movement in structured storage, where reliability and throughput beat learned dexterity. Buy or build the mechanism; treat AI as an optimiser rather than a policy.

Symbotic acquired Walmart's Advanced Systems and Robotics unit for $200M in January 2025 with a commitment to build Walmart's Accelerated Pickup and Delivery centres, hired James Kuffner (ex-TRI CEO, Google robotics co-founder) as CTO the same month — a clear signal it intends to buy into learned manipulation — and acquired autonomous forklift maker Fox Robotics in February 2026.

Mytra offers 3D cellular storage: bots moving inside a lattice, 3,000 lb lift via a "Helix" mechanism, claimed 99.999% uptime, eleven cameras per bot, 150 presentations per hour, slotless inventory.

The key contrast with §6.1–6.8: both generate colossal telemetry, but it is logistics data — throughput, congestion, uptime — not contact-rich manipulation data. Telemetry is not a manipulation corpus.

Skeptic: Symbotic has restated financials and remains customer-concentrated in Walmart; Mytra's uptime and throughput figures are vendor-supplied.

6.11 Manufacturing: Bright Machines and Machina Labs

Bright Machines sells "software-defined manufacturing" — robot cells composed into microfactories — and has repositioned around AI infrastructure assembly, building servers and racks for the AI buildout itself, with a Microsoft Azure partnership. Roughly $400M raised since 2018.

Machina Labs applies two robot arms as an incremental sheet-forming pair, replacing hard tooling: $124M raised in February 2026, a July 2026 Lockheed Martin deal to qualify robotically formed parts for the JASSM missile, and a September 2025 Toyota partnership.

Why Machina's data loop is interesting: each formed part yields a springback and deviation measurement against CAD, so the model learns material response — a physical-property dataset no simulator currently reproduces well. This is a genuinely different data class from every other entry in this chapter.

Skeptic: both are capital-heavy; defence qualification cycles are long, and "software-defined manufacturing" has repeatedly failed to generalise beyond the launch customer's product family.

6.12 Serve Robotics — a natural experiment in whether scale alone buys data

Why it belongs here: sidewalk delivery is the only embodied category where a startup has put thousands of units into uncontrolled outdoor environments. It is the cleanest test of whether deployment scale alone produces a data advantage.

Deployed: by December 2025 Serve had 2,000+ robots deployed via Uber Eats across Los Angeles, Miami, Dallas and Atlanta — a twentyfold fleet expansion during 2025.

The tell: Serve has been acquisitive precisely because fleet scale did not buy it the capability it wanted. It bought Vayu Robotics (September 2025) to combine "Serve's autonomy stack and dataset with Vayu's AI foundation models," bought Voysys' video-streaming assets (August 2025) because teleoperation bandwidth is the hidden cost of remote supervision, and then bought Diligent Robotics (January 2026) to move from navigation into manipulation indoors.

The lesson: 2,000 robots and millions of miles of sidewalk produced navigation and pedestrian-interaction data that transfers poorly to manipulation. Scale is necessary but not sufficient; the right kind of scale is what matters.