Gemini 4 Argon and Pi API beta

Today: Google’s physical reasoning model, a robot foundation model API from Physical Intelligence, a BMW dexterity partnership, a 1,250 home factory deal, ABB’s vision tooling, and automated red teaming for physical AI.

Gemini 4 Argon targets the robotics stack

Google published Gemini 4 Argon for Robotics on 13 October, a general purpose physical reasoning model pitched across four areas: robot control, real to sim reconstruction, physical understanding, and self improvement (Google Developers Blog). Google says Argon scores 87 percent on the Embodied Reasoning benchmark (ERQA), a test of spatial and physical problem solving in 3D environments, and beats every other frontier model using pure reasoning. Treat that as a vendor claim on a vendor benchmark.

The demos matter more than the number. Argon drives a Boston Dynamics Spot (12 degree of freedom legs, 6 degree of freedom arm) to step on a foot pedal bin and throw away soda cans, after building an articulated MuJoCo twin of the bin to practise the move on. A second workflow has the model emit motion primitives such as robot.grasp and robot.move_until_contact against a local HTTP endpoint, while reading a wrist camera image annotated with a 2 cm world coordinate grid. The pattern worth copying is code as policies over a small primitive set, plus practice in simulation. See our VLA model primer and the world action model conditioning piece for background.

Pi API puts robot foundation models behind an endpoint

Physical Intelligence opened a beta of the Pi API on 13 October: upload your data, launch fine tuning on compute the company manages, then run the model on your robot through a hosted inference gateway (Physical Intelligence). The models underneath are the general purpose π0 and π0.5 family, which the company says now fine tunes more data efficiently and performs better for its partners.

Early access partners show how small a slice of the loop a developer still owns. Actor runs an autonomous wheel loader loading a dump truck in a commercial deployment with nobody inside the machine. Almond builds Axol, a dual arm robot, and used a model obtained after four days of API iteration to deburr spacers used in its own assembly. Cloud hosted inference covered most applications shown, and Physical Intelligence names network connectivity as a weakness, which is the argument for keeping a smaller policy on a Jetson class board when the link is unreliable. The loop the API encodes, collect demonstrations, train, test, then use failures to decide what to collect next, is the data bottleneck restated as a service.

Industrial Next takes flexible factory stations to BMW

Industrial Next, a San Francisco startup founded in 2021 by two former Tesla engineers, announced at TechCrunch Disrupt that BMW Group Technology Office will evaluate its AI powered dexterous manipulation for manufacturing (TechCrunch). Its self contained workstations, driven by proprietary foundation models that connect vision, task context and motion, are said to be retoolable in minutes rather than weeks, with new tasks taught through simulation and demonstration instead of hard coded scripts.

Cofounder Lukas Pankau describes one station making a popular smartphone, threading a flexible cable and snapping a connector cap, work he says manufacturers failed to automate for years. A film peeling station is said to strip protective film off camera modules and buttons at or above human speed with 100 percent accuracy, including after moving to a phone model and line it had never seen. The underlying claim is an extremely high fidelity simulation engine, close enough to real dynamics that high precision assembly runs almost without seeing the parts: the sim to real argument taken to its extreme. None of it is independently measured yet.

Promise Robotics signs D.R. Horton for 1,250 homes

Promise Robotics signed a multi year agreement with D.R. Horton to produce more than 1,250 homes in the United States through its Factory as a Service model (The Robot Report). The Toronto company, founded 2021, runs distributed microfactories rather than one large line, and currently uses cloud based simulations with KUKA industrial arms while staying robot agnostic, per CEO Ramtin Attar.

The numbers are manufacturing geometry rather than hype: 45 ft floors, 20 ft walls, and a sweet spot of three to four storey townhomes, because taller buildings need sprinklers, bracing and metalwork. Attar says a two family home with two storeys, a basement and a garage can be produced in one day, excluding electrical work, windows and plumbing, against a typical two to three weeks. Per home the company says it generates 300 new images and 5,000 new tactile sources, using computer vision plus LLMs for data preparation plus a domain specific reasoning model. Distributed microfactories are a friendlier problem for learned policies than one retooled line, and the variation between every wall is what robot data pipelines struggle with.

ABB folds AI vision training into OmniCore

ABB Robotics is expanding OmniCore EyeMotion with AI Robot Trainer, aimed at system integrators who want to deploy vision applications without building models from scratch (Design World). Image acquisition, model training, deployment and model updates sit in one workflow, and automatic collision free path planning converts vision output into robot motion. ABB claims up to 90 percent less commissioning time and up to 50 percent less cycle time with that path planning.

Training blends synthetic and real images, cameras and sensors from multiple vendors are supported, and a guided workflow plus an AI chatbot covers changing SKUs, lighting and layouts. Applications include 2D pick and place, 3D bin picking, inspection, sorting and assembly. The value for integrators is the update path: vision models that drift with lighting and parts are the usual reason a working cell stops working.

Laplace Intelligence red teams robots in simulation

Laplace Intelligence, an Ann Arbor startup founded a year ago out of the University of Michigan, is building automated red teaming for robots, autonomous vehicles and other physical AI systems (TechCrunch). CEO Howie Sun and chief scientist Henry Liu published work on challenging autonomous driving AI in simulation that reached the cover of Nature in 2023, and the company is presenting in the Startup Battlefield 200 at TechCrunch Disrupt this week.

The product pairs a high fidelity simulation reconstructed from real world data, built on a post trained version of NVIDIA Cosmos, with agents that hunt for failure cases instead of waiting for humans to write them. Environments are generated to scale by blending map data with reference imagery, traffic and congestion feeds come from providers such as TomTom, and other actors react interactively rather than replaying scripted trajectories, with scenario seeds mined from news and police collision reports. Verification is the missing half of most safety stories, and it pairs with the gate in front of the model and a guard layer in the stack.

Quick Bites

• PlanSys2 is switching its defaults to Kairos, a temporal and numeric PDDL planner developed inside the project that solves more problems than LAMA, OPTIC, TFD, POPF and ENHSP on the same benchmarks, and to STN for plan behaviour trees, which supports required concurrency (ROS Discourse).
• Perseverance set a Mars driving record on 11 October: a 299.6 ft drive took its total to 28.10 miles (45.22 km), past Opportunity’s 28.06 mile record, helped by years of autonomy upgrades, per JPL’s Mark Maimone (IEEE Spectrum).
• eProsima previewed an MCP server in DDS Monitor Pro that lets an AI assistant explore discovered applications, topics, publishers and subscribers, trace a reliability QoS mismatch and publish test samples. It is commercial, in final testing, and starts disabled (ROS Discourse).
• Gazebo’s PMC reports Ionic reaching end of life in December 2026, Homebrew dropping support for Harmonic and Fortress on 10 January 2027 when FreeImage is removed, and Qt5 leaving Homebrew in May 2027 (Gazebo PMC).