Manufacturing and robotics: what AI is really doing

What's actually changing

1. 1

   Humanoid robots move from demo livestreams to paid warehouse shifts

   How it works. Agility's Digit and Figure's Figure 03 are running inside real warehouses, not just labs: Digit has logged over 65,000 hours at Amazon, Schaeffler, GXO, Toyota and Mercado Libre, and Figure sped its BotQ factory from one robot a day to one an hour. The robots still lean on a mix of scripted routines, teleoperation and AI for the hard parts, and most units shipped so far go to R&D and data collection, not a customer's payroll line.

   Why it matters. Warehouse operators can now point to logged hours and named customers instead of a keynote video, the first real evidence that a leased humanoid might beat hiring for repetitive tote-moving and machine-tending work.

   What's overstatedThe viral demos, Figure's 81-hour livestream chief among them, are R&D and data-collection runs, not sales. Chinese manufacturers are also shipping humanoids in volume while Western firms are still counting pilot hours, so the two markets are not competing on the same numbers yet.

   Figure AIAgility RoboticsAmazonSchaefflerGXOToyota Motor Manufacturing Canada

   3 receipts

   • Figure sped Figure 03 production at its BotQ factory from one robot a day to one per hour, a 24x throughput jump in under 120 days. delivering over 350 of our third generation humanoid robots and increasing our production rate from 1 Figure 03 per day to 1 per hour - a 24x throughput improvement in under 120 days Figure AI, 29 Apr 2026 ↗
   • Agility's Digit has logged more than 65,000 hours of real-world operation at named customers. more than 65,000 hours of real-world operation overall GeekWire, 24 June 2026 ↗
   • Digit is rated to lift 35 pounds and handle repetitive warehouse tasks for up to 20 hours a day. It lifts up to 35 pounds and handles repetitive tasks like moving totes and tending machines for up to 20 hours a day. GeekWire, 24 June 2026 ↗
2. 2

   One AI model to run any robot body: physical-AI foundation models arrive

   How it works. Physical Intelligence's pi-0.7 and Skild AI's Skild Brain both claim a single trained model that can drive different robot bodies, arms, quadrupeds, humanoids, and attempt tasks it never saw in training, like folding laundry or using a new kitchen appliance. Nvidia is trying to standardize the hardware side too, pairing its Isaac GR00T software with a reference humanoid body (Unitree chassis, tactile hands, an onboard Jetson Thor chip) so university labs build on one shared stack instead of ten incompatible ones.

   Why it matters. If a model really transfers to a new robot body without retraining, the cost of programming a robot for a new task or a new factory collapses, the same shift that made large language models cheap to redeploy across products.

   What's overstatedA 'step-change in generalization' claim on a company blog is not a third-party benchmark, and none of these foundation models has a named factory paying for production use yet. They show up in demos and funding announcements, not shift logs.

   Physical IntelligenceSkild AINvidiaUnitree

   3 receipts

   • Physical Intelligence released pi-0.7, which it says generalizes to tasks like laundry folding it never saw in training. We've trained a new model, π 0.7, that exhibits a step-change in generalization. Physical Intelligence, 16 Apr 2026 ↗
   • Skild AI describes the Skild Brain as one foundation model meant to control any robot body without prior knowledge of its form. It is the industry's first unified robotics foundation model that generalizes across tasks and robot hardware. Skild AI, 14 Jan 2026 ↗
   • Nvidia opened a reference humanoid robot design pairing a Unitree chassis with its Isaac GR00T software stack for research labs. The NVIDIA Isaac GR00T Reference Humanoid Robot is a state-of-the-art platform that brings the key building blocks for frontier humanoid research into one system, pairing a human-scale robot body with dexterous manipulation, sensing, control and onboard AI compute. NVIDIA Newsroom, 31 May 2026 ↗
3. 3

   Autonomous trucks drop the safety driver, for paying freight customers

   How it works. Aurora, Kodiak and Gatik are running trucks with nobody behind the wheel on public roads, hauling freight under contract for named shippers (McLane, Atlas Energy, PepsiCo, Loblaw), not just test loops. Waabi went further, moving its driving software from a Peterbilt truck to a Volvo VNL with no retraining, evidence that the model, not the truck, is what it sells.

   Why it matters. Freight beat robotaxis to true driverless operation in most markets because highway lanes and fixed routes are a narrower problem than city streets, making trucking the clearest real-world test of whether removing the driver actually works at scale.

   What's overstatedThe dollar signs are still small: Aurora's entire first-quarter 2026 driverless-plus-supervised revenue was $1 million, next to funding rounds and valuations in the billions. 'Driverless' can also still mean a company support rider in the back seat gathering feedback, not zero humans in the vehicle.

   Aurora InnovationKodiak AIWaabiGatikMcLanePepsiCo

   3 receipts

   • Aurora's trucks for McLane run autonomously with no human safety driver able to take over. These trucks will operate autonomously and will not have a human safety driver on board who can take over. TechCrunch, 6 May 2026 ↗
   • Aurora's own Q1 2026 shareholder letter put driverless-plus-supervised revenue at $1 million for the quarter, alongside a perfect safety record. First quarter 2026 revenue totaled $1 million across driverless and vehicle operator supervised commercial loads Aurora Innovation SEC filing, 6 May 2026 ↗
   • Waabi transferred its driving software from a Peterbilt truck to a Volvo VNL Autonomous with no new engineering or retraining. took control of a Volvo VNL Autonomous truck and drove it safely on highways and complex surface streets from the very first mile FreightWaves, 15 July 2026 ↗
4. 4

   AI turns into production-line software: welding, food assembly, machine health

   How it works. Path Robotics put its Obsidian welding model on a mobile quadruped, Rove, so it walks to a ship hull instead of needing the hull brought to a fixed welding cell. Chef Robotics' arms passed 100 million meal servings across a dozen-plus plants. Augury's own survey of manufacturers found the share running AI across more than half their facilities tripled in a year, from 14% to 42%.

   Why it matters. These are labor-shortage trades, welding, food-line assembly, machine monitoring, where physical AI already has customers writing checks, making this the least glamorous but most commercially real part of the sector.

   What's overstatedPredictive-maintenance vendors routinely market unaudited '99%+ accuracy' numbers, and Augury's own report means 58% of manufacturers are still piloting AI or have not started, even as the adoption headline reads as a tipping point.

   Path RoboticsChef RoboticsAuguryHIIRockwell Automation

   3 receipts

   • Path Robotics launched Rove, mounting its welding AI on a mobile quadruped to reach large, immovable structures like ship hulls. a mobile robotic welding system designed to address the historic difficulty of automating welds on large, immovable structures common in shipbuilding and heavy construction Ohio Tech News, 24 Apr 2026 ↗
   • Chef Robotics' AI-driven arms have produced 100 million meal servings across more than a dozen plants in the US, Canada and Europe. its robots have completed 100 million servings in production at customer facilities The Robot Report, 17 Apr 2026 ↗
   • Augury's report found the share of manufacturers scaling AI across more than half their facilities tripled year over year, from 14% to 42%. The share of organizations scaling AI across more than half their facilities has tripled year-over-year, rising from 14% to 42%. Augury, 9 June 2026 ↗
5. 5

   Supply chain software adds an AI agent instead of just more automation

   How it works. Augment's AI agent, Augie, now runs inside the software freight brokers and distributors already use, automating carrier selection and shipment tracking inside McLeod's PowerBroker TMS and order entry for a distributor's quote-to-cash process. Symbotic, meanwhile, bought a UK warehouse-intelligence firm, ARMS Innovations, to bolt real-time operations software onto its physical robots instead of only selling hardware.

   Why it matters. The money is shifting from selling a robot or a software license to selling an AI teammate that plugs into the transportation and warehouse-management systems already running, which makes the old-line TMS and WMS vendors the distribution channel for AI, not the target it disrupts.

   What's overstatedCalling this an 'AI teammate' oversells the autonomy: these are agentic wrappers automating carrier selection and order entry inside existing workflows. Someone still owns it when a shipment goes wrong.

   AugmentSymboticMcLeod SoftwarePenske LogisticsMedline

   2 receipts

   • Augment's Augie went live inside McLeod Software's PowerBroker TMS, automating carrier selection and shipment tracking for freight brokers. carrier selection and shipment tracking Augment, 24 Aug 2026 ↗
   • Symbotic acquired ARMS Innovations, a UK firm specializing in real-time operational intelligence for automated warehouses. a UK-based software company specializing in real-time operational intelligence solutions for complex automated warehouse environments GlobeNewswire, 2 July 2026 ↗

Who's doing it

14 companies to know, 9 with a round we can source. Newest first where dated.

Search these companies in the startup database ↗ · Explore all research tools

Your move

Our read, built on the receipts above.

What's next

• Dec 2026 Kodiak AI's targeted commercial launch of driverless long-haul trucking kodiak.ai ↗
• Dec 2026 Aurora's year-end target of over 200 driverless trucks running across the Sun Belt sec.gov ↗