Isaac Lab 3.0 goes GA (8,437 stars)

Robot training is moving to simulation first, so the plant's digital model becomes the training ground.

A Spot robot now makes four inspection rounds a day at a Korean power plant

Korea Western Power is using a Boston Dynamics Spot, set up by Korean integrator Clobot, for four 90-minute rounds a day at its Gimpo plant, carrying thermal, ultrasonic, lidar and gas sensors. During validation it caught a natural gas valve anomaly.

Image: Western Power

COPA-DATA puts its zenon SCADA manual behind an MCP server, so an AI assistant can cite it (0 stars)

The server lets an AI assistant search the official zenon documentation and pull full topics with citations, instead of guessing from memory. COPA-DATA says it is not a supported product and anything safety-relevant should be checked against the source docs.

For SCADA and HMI engineers on zenon: grounded answers about the software without digging through the help site.

Antropi Robotics comes out of stealth with an autonomous factory for metal parts

Antropi says it built its first fully autonomous metal parts factory in six months, is shipping, and turned profitable without outside money. Its goal is US quality at China prices, with delivery in under 24 hours.

Anduril's Ghost Shark underwater drones are now coming off an Australian production line

Anduril and the Australian government confirmed the large autonomous sub can carry drone swarms, sea mines and logistics payloads. Australia's defence industry minister says output is rising less than a year after the factory opened.

Browser Use open-sources rl-env-use, which builds training environments from a one-line skill description

You describe a skill and get a verified environment built from real cases. It's an early path to training agents on an operation's own tasks.

Arda, from OpenAI's former research chief, raises $232 million to turn CAD files into factory assembly lines

Arda's agent works from a product's CAD files to set the assembly order, write the work instructions and lay out the stations, and Anduril says basic work instructions dropped from 10 hours to 15 minutes (company figure). After defense, the first markets named are makers of transformers and electrical equipment for the data center supply chain.

Found via @bobmcgrewai on X

Image: Simon Kadula on Unsplash, logo by Arda Industries, via The Next Web

Toyota Material Handling says digital twins can halve commissioning time for autonomous forklifts

Customer sites are rebuilt in NVIDIA Isaac Sim so navigation, safety logic and traffic with people and manual forklifts get tested before a truck arrives, and synthetic images raised pallet detection recall by more than 70%. Toyota also targets up to 40% faster time to market; all figures are company figures.

Carbon: an open-source ERP, MRP, MES and quality system for shops, with its own API and MCP server

Carbon keeps planning, production, quality and accounting in one Postgres database, from quote to invoice, with lot and serial tracing. An agent can read that one schema through the API or MCP server instead of reconciling four systems that never agree.

For fabrication shops and small manufacturers: one open system an agent can query, from the same company that sells Carbon Cloud.

DataSteward: an agent can query plant data and propose a change, but a manager approves it and every step is logged

In the demo, a buyer's agent proposes a safety-stock change and a manager reviews the query, the policy decision and the audit record before it goes through, with rollback for supported changes. It is a reference project on synthetic data, and the README says the approve step alone does not prove a person approved it.

For IT and operations leads deciding how much an agent may touch inventory and planning data: preview, approve, audit and roll back is the pattern to ask for.

new-acad: an AI assistant inside AutoCAD and Civil 3D with 270 MCP tools for alignments, corridors, pipe networks and grading

It adds an AI palette inside the app, 270 MCP tools and more than 390 plugin methods for surfaces, alignments, corridors, pipe networks and grading. The README is Chinese first with an English version, and any OpenAI-compatible key works.

For civil designers working site grading, roads and utility pipe networks in Civil 3D: try it on a copy of a drawing first.

Pascal Editor: a free 3D building editor that AI agents can drive through MCP, 25,000 GitHub stars

It is an open-source 3D architectural editor with a local CLI and MCP tools. The team says it is exploring facility management and infrastructure, and its hosted preview shows the structure and systems behind the walls, which is not yet in the open-source release.

For facility teams and owner's reps: quick walkable building models without a full BIM seat.

Goldman counts 86 data center projects canceled or significantly delayed since January, 76 of them over local opposition

The report, shared with Axios, also puts the median time from a grid connection application to commercial operation at roughly five years. The bottleneck for the buildout is permits, neighbors and power, not chips.

Google's 780 MW Pine Island data center comes with 1,900 MW of new wind, solar and 100-hour storage, and Google pays for the grid work

Xcel says Google will cover any new grid infrastructure, and the storage piece is a 300 MW, 30 GWh Form Energy iron-air battery. The deal includes minimum payments and an exit fee if demand falls short, and it is still waiting on Minnesota regulators.

An oil and gas data engineer turned 12,000 scanned land and title documents into one graph before letting an agent near them

Every well is tied to its lease, every lease to its tracts, and every answer points back to the page it came from. His line: solve the data problem first, because the agent is easy and trust is the hard part.

Microsoft walked away from a 50 MW UK data center because grid power will not arrive until the 2030s.

Microsoft dropped its anchor deal for Nscale's planned 50 MW site in Loughton, Essex, after warnings the grid might not deliver power until the early to mid 2030s. The plan was 23,040 Nvidia GB300 GPUs. The customer, chips and money were in place, and the interconnection date still decided it.

Found via Kyler Schmitz on LinkedIn

CMMS MCP server, 0 stars: an AI assistant looks up assets, acknowledges alarms and moves work orders, with role checks

Work orders follow a real life cycle with an audit trail, and a read-only mode leaves the write tools out entirely. It ships with a made-up plant and a companion skill that teaches the agent the safe order of calls.

For maintenance planners and reliability engineers: a working pattern for letting an agent touch work orders without giving it free rein.

Motive launches agents that open maintenance work orders and call drivers before a breakdown

The agents use data already in Motive's fleet platform to contact drivers, create work orders and alert managers without someone working every alert by hand. Motive has not released independently checked results, and its savings figures are vendor-reported.

Open Industry Project: a free browser-based factory simulator with conveyors, sensors, robots and its own PLC

The built-in PLC runs a subset of Structured Text and Ladder Diagram, and a desktop app connects a real PLC over Modbus TCP, OPC UA, EtherNet/IP, S7 or ADS. The team says it is in early development and breaking changes can land in any update.

For controls engineers and trainers: test PLC logic on a virtual palletizing line without buying hardware or PLC software.

Image: Open Industry Project

Graph Overlay Inspector: lay a P&ID's extracted knowledge graph on top of the drawing and check it tag by tag

You inspect the graph on the drawing it came from, run SPARQL with the results shown on the sheet, and send fixes back into the graph through a change log. It runs as a Claude artifact or a single HTML file, and Claude features bill the viewer's own account or API key.

For engineers digitizing P&IDs: see where an AI extraction is wrong before anyone builds on it.

Awesome AI for 3D modeling and robotics: more than 700 source-linked cases of AI doing CAD, BIM drafting and robot control

Each case links the original post, names the author, date and model, and quotes the author for any number, and the demo videos are archived in the repo. It is the companion to a survey from MIT CDFG.

For engineering managers: see what AI has actually done in parametric CAD and robot control, with the model named on every case.

A power plant operator on 12-hour shifts who had never written code built his crew's shift planner in about 8 hours

It has a rotation calendar, call-ins, overtime tracking, time off and paycheck estimates, and his crew was already asking for features. He built it by talking to Grok Bot and Cursor, a sign that workforce tools are starting to come from the workforce.

Hitachi Energy's CEO says power equipment supply and demand are out of balance, and India is now central to its supply chain

Andreas Schierenbeck wants customers to sign framework agreements and reserve capacity years ahead, since a data center can be built in 1 to 3 years while the generation and transmission to serve it take about 10. Hitachi Energy began building a large power transformer factory in Vadodara in July.

Bagel: ask robot, drone and sensor logs questions in plain English, and see the SQL it ran

Every calculation is a DuckDB SQL query that Bagel shows you, so the answer can be audited. The README says it keeps the model in front of your logs and never in the robot's control loop.

For inspection and robotics teams sitting on drone, crawler or IoT recordings: get answers without writing queries, and check the query when it matters.

RAX: a coding agent learns a robot task by trial and error, then hands back a program that runs without the agent

A coding agent such as Codex CLI or Claude Code drives the robot, learns the task from its own tries and compiles what worked into a robot program. The agent is the programmer, not the operator, and plain code is something a plant can review.

For automation engineers and integrators who program robot cells: what you validate is ordinary code, not a live agent.

Cad2earth: a Claude Code skill that turns AutoCAD drawings into a clean Google Earth file

It finds the coordinate system Civil 3D embedded, keeps CAD layers as folders, turns blocks like valves and pumps into symbols, and prints a report of everything it skipped and why. That report is the honest part.

For land, right-of-way and survey teams: put design DWGs in front of managers in Google Earth, and read the skipped list before you share it.

Siemens and NVIDIA: an AI agent drafts production plans at the Erlangen factory and a digital twin checks them before anything changes

When new orders arrive, the agent generates production scenarios and shows planners the impact of each, and a digital twin of the whole plant validates them before they are used. Siemens claims up to 20% cost savings from the planning agent and up to 85% less engineering effort for virtual ramp-up of new circuit board assemblies.

Image: Siemens

Hyundai and Kia start their Ulsan EV plant and show Polaris, a plant AI that answers "show me production status" in a second

Polaris, in pilot at the Ulsan plant since September, answered within one second with operating rates, body defects and suggested fixes. The bigger move is one standard manufacturing platform across plants; humanoid robots are on the roadmap from 2032, not now.

Image: Electronic Times

ISU Chemical gets an AI that recommends setpoints for its continuous units, but operators approve every change

Kolon Benit's system will use time-series AI to spot early signs of process trouble and recommend operating conditions, and setpoints reach the control system only after operators approve them within safe limits. That is advisory control, the version most chemical plants can accept.

Image: Kolon Benit, via Chosunbiz

Amazon, AEP and X-energy plan a 960 MW small reactor plant on a retired Indiana coal site

The plan is twelve 80 MW Xe-100 units at the former Breed station in Sullivan County, with Amazon funding development and buying the power. It is still a feasibility deal with no final investment decision, and X-energy has yet to break ground on its first plant in Texas.

Cloudflare's Clairic pulls typed numbers, dates and values out of messy text in one pass

Kevin Flansburg says Clairic is built on regression instead of classification, so it can estimate values without a list of choices, and open weights are coming. Cloudflare's examples return each value with a confidence score, or null when it is not there, which suits invoices, delivery records and repair estimates.

Sunra MCP puts image, video and audio models inside Claude and ChatGPT at their makers' API prices

Sunra says every leading image, video and audio model is available inside the assistant you already use, each at its maker's own API price with no markup. That is one connection instead of a separate account per model.

For marketing and sales teams making product images and video: compare the per-model prices with what you pay now.

SKY-001, an engineering world model, plans CAD and BIM designs, and its makers claim it beats frontier models

Sam Pasupalak says it is 51 to 125 times faster and 150 to 210 times cheaper than GPT-6 Astra on CAD and BIM tasks, with higher IOU than every frontier model they tested. Those are the company's own benchmarks, so test it on your own parts.

Cloudflare open-sourced the security audit skill behind its own vulnerability hunting

It runs six phases, from mapping the codebase to hunting area by area, and every finding goes to a fresh agent that did not find it to confirm it is real. You get a findings file split into confirmed, needs validation and rejected.

For teams writing plant software, scripts or integrations: run it on your own code and fix the confirmed findings first.

Viggle releases JST-2, a world model built in 3D instead of video

Viggle says JST-2 generates a 3D scene in seconds, expands it in real time and creates characters with motion built in, all in one architecture. Models that work in 3D rather than pixels are closer to what plant layout and training scenes need.

Google used Gemini 4 Argon to turn a photo of a desk into a playable 3D track

Argon orchestrated subagents, Blender MCP and Nano Banana to turn everyday objects on a real desk into the track. The same photo-to-3D chain is a cheap way to rough out a real space before modeling it properly.

IKEA's real BILLY manual, rebuilt in 3D so you can watch every step as drawn

Ryan Sael's version follows the manual page for page, with the same parts, hardware counts and order, and he notes panel thickness and hole positions are read off the drawings, not measured. Work instructions for plant equipment could be shown the same way.

An agent tree component shows which AI agent started which

Flo shared a component that draws agents as a tree, so you can see who started whom. As agents start other agents, that kind of view is how a person keeps track of what is running.

A maintenance agent for alarm triage got the cause, priority and parts right on 17 of 20 alarms, and its author lists the 3 misses

It triages alarms on CNC spindles and air compressors, checks the storeroom and drafts the work order, with no AI model call in the alarm path. The misses are the useful part: one was a failed sensor reading 131 C that the rules read as cooler fouling.

For maintenance planners and reliability engineers: copy the method, which is test the agent on real alarms and write down where it was wrong.

SemiAnalysis forecasts 51.2 GW of new data center capacity outside China in 2027, double this year

It points to more than 24 GW already in advanced construction at 40 campuses for 2027 delivery. For people who build, power and equip these sites, the forecast reads as another year of long waits for transformers, switchgear and crews.

PowerMCP: an AI assistant that runs real grid studies in PowerWorld, PSS/E, OpenDSS and pandapower

The solver does the math and the agent only drives it: it sets up cases, runs power flow and contingency checks, and drafts the report. Its skills pack will not run advanced studies until the base case solves.

For transmission planners and interconnection engineers: hand the case setup and the write-up to an agent and keep the engineering sign-off.

ScottMadden counts more than 60 AI deployments in U.S. nuclear plants, and about 60% have no measured benefit yet

Nuclear already uses AI; the gap is proof. One operator built an AI observation tool into a required process and saw it used fleetwide, while an optional AI review tool got uneven use.

Harness²: Google Research open-sources agents that tune their own setup

Google Research open-sourced Harness², which lets AI agents test and improve their own prompts, tools, memory and workflows while the model stays the same. For operations teams, that points to agents that get better at a repeat job without a model swap.

Loböle packs open CAD, FEM, CAM and an MCP server in one Linux tool (5 stars).

Built on OpenCASCADE, it covers sketches, assemblies, drawings, static and modal FEM, topology optimization and milling output for Mach3, LinuxCNC or GRBL, all reachable from Python and over MCP. By the author's own warning, the CAM has cut real parts on only one machine.

For small machine shops and design engineers: worth a test, not a production CAM yet.

How to test subsurface agents without overfitting to a few fields

Agents tuned on three famous fields will look brilliant until they meet a fourth.

IFI Gatehouse, 0 stars: a fence contractor's estimating tool where AI reads the bid package and a named person approves every change

The AI proposes findings from drawings, specs and addenda, each tied to the sheet, page and quoted words, and nothing enters the scope until a person accepts it. The motto is AI interprets, code calculates, humans authorize.

For estimators and preconstruction managers: a concrete way to use AI on bid documents without letting it fill gaps with confident numbers.

Why does AI code like a human? An analytical chemist asks for AI that runs offline and gives the same answer every time

By his own measurements, more than half of what Claude did in a session was drift away from what he asked for. His point for plants: refineries and labs will not hand their data to a model in someone else's cloud, so AI they trust has to run on a laptop, air-gapped, and give the same answer every time.

Image: Medrado Michael Leal, Pathfinding

Industrial Agent Evaluation Lab: replayable tests for maintenance agents, with fault injection and scoring

The agent has to work out which pump readings and maintenance revision to trust, find a qualified technician, reserve the right part, ask for supervisor review and draft a work order. A deterministic evaluator grades its tool calls and the final database state, with injected timeouts, rate limits, duplicate deliveries and a lost work-order response.

For reliability teams and vendors: prove a maintenance agent recovers from bad data and dropped responses, not just the happy path.

opcua-mcp: ask an OPC UA server questions in plain English, read only

An agent can browse the node tree, search for tags like "Flow" and read live values, which is the slow part of a brownfield assessment. It ships with a water treatment plant simulator to practice on.

For SCADA and integration engineers: map what a server exposes in minutes before you build anything on top of it.

PyPSA: the free grid planning tool used to size solar, storage and transmission

Open-source Python maintained at TU Berlin that runs dispatch and capacity planning, with more than 2,200 GitHub stars. Because it is plain Python, an agent can drive it to test how much battery a site really needs.

For power and data center developers: size onsite generation and storage with the same tool grid planners use.