Good morning, robotics enthusiasts. Amazon's newest warehouse robot doesn't need a programmer, just a plain sentence telling it what to move.
The next-gen Proteus is part of a more than €10B push to modernize Amazon's European warehouses. When anyone on the floor can direct a robot in plain language, how fast do warehouses change?
In today's recap:
Amazon's Proteus learns to take plain-language orders
Nvidia and Unitree open a humanoid platform to labs
Generalist AI raises $400M for physical AGI
AMAZON
Amazon's new Proteus robot takes plain-language orders
Recaply: Amazon just unveiled a next-generation Proteus robot that workers direct with plain, conversational text prompts. It's part of a more than €10B plan to expand and modernize Amazon's European fulfillment centers.
Key details:
Workers give Proteus plain-text instructions instead of technical commands, and it works out the priority, route and timing itself. It can now operate anywhere on a site, not just dock areas.
Today's Proteus works at 25 US sites, moving carts of up to nearly 400kg. Amazon also plans to add 25,000 European fulfillment jobs and bring its STARK tote robot to 15 sites.
"You tell it what needs to be done. It figures out the priority, the route, the timing," Amazon Robotics VP Scott Dresser told Reuters.
Amazon showed the robot at its Delivering the Future event in Dartford, England, on June 4. It's in pilots at Amazon's labs, with European deployment planned for the first half of 2027.
Why it matters: Directing warehouse robots has usually meant technical commands or programming tools. Letting floor staff give orders in plain language could make robots far easier to use across Amazon's 230,000-person European workforce. Amazon says the robots take on heavy, repetitive work so people can focus on higher-skilled jobs. Proteus is still in pilots, though, and Europe won't see it until 2027.
NVIDIA & UNITREE
Nvidia and Unitree open a humanoid research platform
Recaply: Nvidia and Unitree just opened up a full humanoid research platform, the Isaac GR00T Reference Humanoid Robot. Unitree will sell the hardware as the H2 Plus, with Nvidia's onboard computer and open software built in.
Key details:
Labs get one integrated system instead of a patchwork. It covers data capture with Isaac Teleop, training in Isaac Sim and Isaac Lab, GR00T foundation models and Isaac ROS for deployment.
Its Unitree H2 body has 31 degrees of freedom, and two Sharpa Wave hands bring the total to 75. A Jetson AGX Thor T5000 delivers 2,070 FP4 teraflops with 128GB of memory.
Researchers keep control of their robot data, training data, telemetry and logs, Unitree said in its own announcement of the H2 Plus.
Nvidia unveiled the platform at GTC Taipei on May 31, and Unitree followed on June 1. It'll be available from Unitree in late 2026, with a Unitree G1 workflow coming sooner.
Why it matters: Researchers face a fragmented process spanning hardware, data collection, simulation, training and deployment, Nvidia says. A shared reference robot lets labs like Ai2, ETH Zurich, Stanford and UC San Diego compare results on the same machine. Robotics moves fastest when researchers can "share code and test ideas on real machines," says Stanford's Steve Cousins. The catch is timing, since the hardware won't ship until late 2026.
GENERALIST AI
Generalist AI raises $400M to chase physical AGI
Recaply: Generalist AI just raised $400M at a valuation of about $2B, with Radical Ventures leading the round. It's building foundation models for what it calls physical AGI, AI that works through robots.
Key details:
Generalist trains its models on large amounts of real-world robot data. Its GEN-0 model showed scaling laws in robotics, where bigger models and more physical experience predictably produce more capable systems.
Generalist's total funding now tops $500M. Its GEN-1 model, released in April, hit 99% reliability across dexterous tasks and ran up to three times faster than earlier systems.
Nvidia, Bezos Expeditions and Spark Capital invested again, while new angels include Fei-Fei Li and Naval Ravikant, the company said.
Generalist announced the round on June 4, two months after GEN-1 and about seven months after GEN-0. The money goes to models, data collection, compute and deployments.
Why it matters: Generalist is betting robots will follow the same path as language models: more data, bigger models, better results. It says a flywheel is forming, where better models do more useful work and that work feeds the next models. The pitch is one intelligence that works across humanoids, factory arms and warehouse robots. Proving that in real deployments is the hard part, and $400M buys time to try.
NEWS
Beyond the big stories in robotics today
eternal.ag deployed its fully autonomous Harvester robot at Dutch grower Van Noord Growers under a long-term deal. The truss tomato picker can run up to 22 hours a day, and Van Noord plans to grow its fleet to three this summer.
Kirisense won Henry Royce Institute funding to build robotic fingertips that sense shear forces and slip in real time, working with the University of Sheffield. The project starts in July 2026 and uses compact optical sensing rather than cameras.
Volvo Autonomous Solutions completed its first autonomous haulage project with Boliden at Sweden's Garpenberg mine. Its self-driving trucks moved close to 700,000 tonnes of rock to reinforce a dam, running over 11,000 cycles and 56,000km.
Neolix partnered with Singapore's QuikBot to build an autonomous delivery chain from public roads into buildings and up to individual doors. The deal supports Neolix's Singapore pilot and targets Asia-Pacific and Middle East markets.
Serve Robotics launched a laundry delivery pilot with NoScrubs in select Los Angeles neighborhoods, its first commercial partnership outside prepared food. The same sidewalk robots that deliver dinner will bring laundry orders to customers' doors.
ABB's B&R division and Salzburg University of Applied Sciences filed a joint patent on AI-based, energy-saving motion control for robots and production lines. A reinforcement learning agent learns which motion profiles waste energy on the real machine.
Robotiq launched IQ, an AI platform that helps partners generate validated robot workcell designs from customer inputs and data from thousands of past installations. CEO Samuel Bouchard says automation "does not scale when integration remains manual."
Fanuc released the CRX-3iA in Europe, its lightest and smallest cobot at 11kg, light enough to carry in one hand. It targets welding in shipbuilding and steel construction, where Europe is short of skilled welders.
Fei-Fei Li's World Labs published "A Functional Taxonomy of World Models," arguing the term has become overloaded. It splits world models into renderers, simulators and planners, and calls simulators the most consequential piece.
Wuji debuted its Wuji Hand 2 at ICRA 2026 in Vienna, a dexterous robot hand with 20 degrees of freedom and backdrivable actuation. The company says an improved thermal design lets it run for longer stretches.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason






