Moving Measurement and Simulation Upstream, and Letting Agents Into Real Devices

This briefing covers September 16–17, 2026 (Beijing time) across 8 sources, spanning AI × industrial design, the latest AI projects, and interesting projects on GitHub. One thread today is about pulling measurement and simulation upstream: a pocket CMF scanner that reads a surface instead of photographing it, and thermomechanical simulation folded into build preparation for metal printing. The other is about agents reaching into real devices — Google opening its smart-home layer as an MCP server, a 9B open-source model that holds its own on spatial reasoning, and a phone that ships "agent" as part of its product definition.

AI × Industrial Design#

1. From "Phone Photos and Memory" to an Archive of Measurements: TRA:CE Reads the Surface Directly(Yanko Design, 2026-09-16; concept by Seongwook Jin): TRA:CE is a coin-sized pocket CMF scanner with a circular display on top and an optical scanner built into the bottom face, reducing the interaction to two moves: set it on the material, press the side button. Instead of reading reflected light, it reads the surface itself, capturing material type, texture, and finish characteristics in a single pass. Each scan becomes a reference card in a companion app, broken into HEX, RGB, LAB, and CMYK values plus the nearest Pantone or RAL match, along with material grain, pore size, and a numeric gloss reading. Why it matters: the least reliable link in CMF communication has always been "photograph a swatch in a showroom, then reproduce it from memory back at the studio." Swapping "guess the reflected light with a camera" for "measure the surface" turns a design team's color and texture references into one reusable, checkable dataset. For teams running material libraries, CMF reviews, and cross-plant handovers, the value of a tool like this isn't extreme precision — it's that the same swatch is still citable two weeks later.
2. Meter-Class Printers Start Changing Tools, Feeding Pellets, and Printing at an Angle: Modix Brings Duet 3 and CANBUS to Its BIG Series(3D Printing Industry, 2026-09-16): Modix has announced Generation 5, moving its BIG series of large-format printers onto the Duet 3 control platform with CANBUS distributed electronics, expanding control to more than 20 axes and drivers. That headroom enables three new capabilities: automatic tool changing via Bondtech's INDX platform, pellet extrusion, and 45-degree printing. Machines ordered since the summer are already shipping on the new platform, and the company plans to demonstrate the technologies at Formnext 2026. Why it matters: large-format FDM used to sell on size alone, at the cost of a single nozzle, a single material, and hard-to-control support and warping. Tool changing lets one machine combine materials and colors in the same part, pellets cut material cost, and 45-degree printing changes where supports go and how loads travel through the layers. For teams building furniture, props, enclosures, and tooling, those three are exactly what move a large-format machine from "big prototypes" toward "deliverable parts."
3. Putting Thermomechanical Simulation Into Build Preparation: AMCM Integrates PanOptimization's PanX Into EOSPRINT(VoxelMatters, 2026-09-17; AMCM and PanOptimization): AMCM has integrated its EOSPRINT build-preparation software with PanX, PanOptimization's thermomechanical simulation tool, spanning EOS printers from the M 290 class up to AMCM's large-format M 4K and M 8K systems. With the integration, EOSPRINT reads an openjz file into PanX, extracting the geometry, build plate layout, and process parameters a simulation needs, so overheating and thermally induced distortion are assessed before production. PanX still runs standalone with any metal system, but both companies stress that bringing simulation into the preparation workflow is the point. Why it matters: simulation has long been a pre-print "expert step" that means opening separate software and doing another round of manual handoff. Folding it into build preparation makes thermal management a default input to geometry and placement decisions rather than a matter of experience. For metal AM design teams, the "can this print" question moves earlier into modeling — supports, orientation, and allowances have to be planned around thermal history from the start.
4. Custom Orthotics Without Manual Scheduling: Mosaic Launches the Orion Belt-Based Continuous Production Platform(3D Printing Industry, 2026-09-16; Canada's Mosaic Manufacturing): Orion is a production platform purpose-built for custom orthotics in orthotics and prosthetics labs and clinics. Its belt-based architecture moves patient-specific devices through production sequentially with automated queue management, enabling unattended overnight and weekend runs. Mosaic says a single printer can produce up to 300 pairs of orthotics per month, with a standard reference device printed in roughly one hour and 45 minutes and material costs as low as about US$6 per pair. Customers can manage scanning, design, and production through the Stryde software suite, or connect the printer to an existing CAD/CAM setup. Why it matters: the real barrier to "going digital" in custom devices isn't buying a different printer — it's connecting scanning, design, and scheduling into a line that can run unattended. For teams working on wearables, medical devices, and mass customization, these numbers provide a comparable reference point: once per-unit material cost falls to single-digit dollars and queuing is automated away, the remaining gap between custom and mass-produced parts is mostly the design process itself.
5. Building an AI Formulation Database for Printed Pills: University of Mississippi and Colorcon Turn Excipient Compatibility Into a Queryable Reference(VoxelMatters, 2026-09-16): A team led by Mo Maniruzzaman, chair and professor of pharmaceutics and drug delivery at the University of Mississippi, worked with Colorcon to build an AI-powered database meant to help pharmacies 3D print medications more efficiently. The core of the project is pairing active pharmaceutical ingredients with compatible excipients — binders, fillers, and colorants — into what the researchers describe as a reference book for drug manufacturing, so a pharmacist can pick compatible materials and direct a printer to produce customized doses, such as smaller doses for children or timed-release capsules. The team says on-demand printing could save millions. Why it matters: this is a classic sequencing problem — a queryable materials database comes first, stable printing second. Most design teams still keep CMF and consumable knowledge in personal experience and scattered notes, which is why every material change triggers another round of trial and error. Turning "material–process compatibility" into a structured, searchable library is exactly the kind of tedious work AI can accelerate.
6. The Point of IMTS 2026 Isn't How Strong the Printer Is — It's That Additive Got Placed Next to Subtractive(VoxelMatters, 2026-09-17): IMTS made a telling floor-plan change this year, placing additive manufacturing beside metal removal in the South Building instead of on its own island. Across the show, AI, robotics, connected production, and automation appear in nearly every sector; in the Emerging Technology Center, an automated cell combines polymer and wire-arc additive, machining, robotics, sensors, and MTConnect to manufacture drones on the show floor. Italian manufacturer Caracol skipped the AM pavilion entirely, positioning its large-format robotic systems in the automation area so visitors could more easily understand how they fit an existing supply chain. Why it matters: after two decades of selling itself as a disruptive alternative, additive is now answering a different question — where does it sit in my factory? For design teams, that shift lands as very concrete delivery requirements: how datums hand off to subsequent machining, how much allowance to leave, and where post-processing and inspection sit in the sequence. Getting those seams right matters more to whether a design ships than printing a beautiful single part.

Latest AI Projects#

1. Anthropic Merges Claude Chat and Cowork Into One Interface, With Claude Design Working Anywhere in the Window(#product #interface; TechCrunch, 2026-09-16): Anthropic is merging its front ends for Claude chat and Cowork so users no longer have to decide which tab a task belongs in. The unified interface brings chat, Cowork, and the Artifacts interactive workspace into one window, and Claude Design — introduced in April for website and prototype design — works anywhere within Claude. Why it matters: split interfaces have been a persistent friction point for teams adopting AI, because conversation, collaboration, and the resulting artifact live in different places and the workflow has to be carried between tabs. Collapsing all three into one window is effectively an admission that design, prototyping, and engineering collaboration are stages of the same work. For teams choosing tools, interface convergence usually means shared context and shared artifacts, which affects day-to-day speed more than any single capability does.
2. Google Home Opens an MCP Server: Claude, ChatGPT, and Other Agents Can Control Your Devices(#product #protocol; TechCrunch, 2026-09-16; Google opened early access on Tuesday): Google has rolled out early access to a Model Context Protocol (MCP) server for its Google Home ecosystem, letting any MCP-supporting AI agent — the report names Claude, Hermes, OpenClaw, ChatGPT, and Google Antigravity — securely work with smart home devices and access event history. People can use natural-language instructions to review camera summaries, monitor activity, control connected devices, and build their own smart home dashboards. Why it matters: once the control layer of a smart home becomes a protocol interface, products gain a non-human user: the agent. Interaction design now has to answer new questions — how device state is described structurally, which actions an agent may take directly, and which must be confirmed by a person. For teams building appliances and control panels, protocols like this will increasingly be part of the product definition rather than integration work at the end.
3. Google Releases Gemini 3.8 Live and Live Extended Thinking: Voice Agents That Reason and Call Tools Mid-Conversation(#newmodel #voice; MarkTechPost, 2026-09-15/16; released by Google, available today in the Gemini Live API and AI Studio): Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are native speech-to-speech models built for real-time voice agents, positioned as an alternative to cascaded ASR + LLM + TTS pipelines. Live targets scale and cost efficiency while combining conversational ability with visual grounding; Extended Thinking adds multi-step reasoning while it speaks, for high-complexity tasks. Extended Thinking takes the top spot on Artificial Analysis' Speech-to-Speech Quality Index at 82.6, scores 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking, and 97.7% on Big Bench Audio. Both are hosted models with no self-hosted option. Why it matters: voice is shifting from a channel to an interface that gets things done, and the metric shifts with it — from "can it hear me clearly" to "can it complete tool calls without breaking the conversation." For teams building interaction hardware, in-car systems, and appliance voice entry points, latency, interruption handling, and task success rates deserve to be in the acceptance criteria more than voice timbre does.
4. Zidong Taichu Open-Sources ZDTaichu5.0-9B: Eight Firsts Across Nine Spatial Benchmarks, Without Sacrificing General Ability(#opensource #embodiedAI; QbitAI, 2026-09-16; open-sourced by Zidong Taichu): ZDTaichu5.0-9B is an open-source general multimodal model that the team says takes first place in eight of nine authoritative spatial understanding benchmarks, raising the ceiling for spatial and embodied ability among sub-10B general models. In demonstrations it completes tasks like opening a drawer, placing a utility knife inside, and closing it again — a sequence requiring continuous spatial judgment and high-level task planning — while keeping image-text understanding, OCR, math reasoning, and coding in the top tier. The article argues earlier multimodal models either fell short in the physical world or gave up general ability to gain spatial skill, and this release tries to hold both. Why it matters: spatial understanding has long been the trade-off point between embodied and general models, and one 9B model covering both means desktop-class compute can handle perception and planning at once. For teams designing robots and human-robot interaction, that changes how compute and sensor budgets get allocated: if planning runs on a small local model, more of the body's volume can go to thermal design and structure.
5. Jensen Huang at Dreamforce: Safety Is an Engineering Problem, Not a Legal One(#industry #safety; TechCrunch, 2026-09-15): Speaking at Salesforce's Dreamforce conference, Nvidia founder and CEO Jensen Huang made his position clear: AI is not some new form of alien mind but hardware and software built by humans, and therefore controllable by humans and existing laws. "Safety is an engineering problem, not a legal one," he said, on the grounds that these systems are ultimately computing systems. Why it matters: this stands in direct contrast to the frontier labs' push over the past week for a slower pace plus third-party auditing, showing the industry has not settled on a single narrative about safety. For teams selecting models and AI tools, that disagreement will show up in procurement checklists: compliance requirements may come from legislation, from an industry standards body, or from a vendor's own engineering commitments — and those three are verifiable in very different ways.
6. Nubia Launches the NaviX Ultra Agent Phone: Snapdragon 8 Elite Gen 5 Plus the Consumer Edition of Doubao's Phone Assistant(#product #devices; QbitAI, 2026-09-16; announced by Nubia with Qualcomm and Doubao): Nubia has launched the NaviX Ultra agent phone, built on the fifth-generation Snapdragon 8 Elite platform — a custom third-generation Qualcomm Oryon CPU, an Adreno GPU, and a Hexagon NPU tuned for agentic AI — paired with the consumer edition of Doubao's phone assistant. Users can wake the model through several methods and get fast responses across continuous conversation, multitasking, and cross-app operations. Why it matters: writing "agent" into the product definition shifts a phone's selling point from a spec sheet to an always-available assistant experience, which requires hardware teams to plan compute, power, and thermal budgets up front and interaction teams to redesign wake methods and permission prompts. For teams in consumer electronics and smart devices, this is a concrete case of product definition and platform capability pulling on each other.

Interesting GitHub Projects#

1. zhbi98/pcb-layout-design: PCB Layout and Board-Release Prep as a Codex Skill(#opensource #hardware; GitHub, created 2026-09-13, updated 2026-09-14, ~79★, topics include ai-agents, pcb-layout, skills): Aimed at the stage after schematic capture and footprint assignment, it covers layout, routing, copper pours, design review, and board-release preparation. The entry point is a SKILL.md, and the agent pulls the topic rules it needs from references/ based on the task; it can live directly inside a KiCad project directory with no prior installation. Why it matters: it turns the most experience-dependent and hardest-to-restate stretch of hardware design into a rule library an agent can draw on. For small teams handling both industrial design and circuit work, that lands much closer to "reviewable engineering documentation" than everyone writing their own prompts.
2. alchaincyf/3d-vibe-coding-handbook: A 386-Page Field Manual for Image-to-3D, With Six Playable Demos(#opensource #guide; GitHub, created 2026-09-13, updated 2026-09-15, HTML, ~172★): The companion repository for the 3D Vibe Coding Handbook, collecting the HTML edition of the book, demo source code, tooling scripts, and a resource index. It walks the full GPT-6 Astra × Tripo pipeline: getting the first image-to-3D run going in ten minutes, what to do after generation, decimation and retopology, rigging and animation, and how to spend a credit budget. The author claims 386 pages, 256,000 characters, and 260 images — with the failures left in. Why it matters: most public material on image-to-3D stops at "the output looks good," while the part that actually blocks people is the long chain of work that follows generation. Organizing decimation, retopology, rigging, and budget decisions as symptom-to-solution entries makes this a rare roadmap with documented mistakes for a designer wiring generative models into a real project for the first time.
3. haplollc/Minted: Turn Any SVG Into a 3D Gold Medallion With Real Lighting(#opensource #rendering; GitHub, created 2026-09-09, updated 2026-09-11, Swift, ~100★, MIT; iOS 17+/Swift 5.9+): A SwiftUI library that strikes any SVG into a 3D gold coin, with cloisonné enamel inlays, engraved lettering, an orange-peel back surface, and momentum-driven spins — and no 3D asset files required. The author's framing is that achievement badges, collectibles, loyalty stamps, and awards all want to feel like things. Why it matters: it demonstrates a very practical division of labor — designers maintain flat vector assets only, while lighting, materials, and motion are generated in code. For teams building in-app badges, collectibles, and physical-feeling visuals, that is far more realistic than sending designers through a 3D pipeline, and it keeps assets consistent across iterations.
4. wildcard1719/RCWS_v5: A 3D-Printed Remote Weapon Station Prototype With STEP Files and a Full Mechanical Design Handbook(#opensource #mechanicaldesign; GitHub, created 2026-09-11, updated 2026-09-11, ~107★, 22 forks): Publishes the STEP models for the RCWS v5 prototype along with a Mechanical Design Handbook whose table of contents covers system architecture, the drive system (main frame and neck frame, main actuator, 3D-printed azimuth bearing, tripod), the launcher (feeder and accelerator), the EOTS (actuator and camera frame), and electronic control. Why it matters: open-source projects usually hand over results, not the design process. This handbook lays out subsystems, drive trains, and bearing solutions for a complex electromechanical product, making it directly comparable reference material for teams working on gimbals, robot joints, and motion structures. (It is a weapon-oriented project; what is worth taking from it is the mechanical design reasoning.)
5. sixtysevenlf/dsh-blender-plugin: Let an AI Model Drive Blender Over a Single TCP Channel(#opensource #3Dtools; GitHub, created 2026-09-13, updated 2026-09-16, Python, ~46★, BSD-3-Clause; version 0.8.0): No screenshots into prompts and no MCP server — instead one direct TCP channel gives the model ten primitives: see the viewport, edit the scene, watch over time, run an inner search loop, profile and fix render performance, decimate objects safely, offload heavy work to a headless process, and operate the channel itself. Version 0.8.0 defaults to EEVEE plus ray tracing, which the project measures at 1.35 seconds per warm frame against 3.13 seconds for Cycles GPU; it also documents real pitfalls such as headless Cycles silently falling back to CPU for a 15.4× slowdown. Why it matters: wiring a renderer into an environment the model can repeatedly observe and iterate on is a precondition for AI taking part in appearance decisions — a single static screenshot gives it no way to tell whether a change actually improved anything. The quantified performance regressions are just as valuable, because silent fallback is the easiest risk to miss when connecting AI to a render pipeline.