Hey there, fellow future-addicts!
Welcome to this week's edition of Rushing Robotics! This week's lineup is one of those rare collections where almost every story reads like a headline from a distant future — except it all actually happened. From AI teaching bionic eyes to see to a robot earning its keep on a live Mercedes-Benz production line, the pace of progress across AI, robotics and biotech has rarely felt this tangible.
🤯 Mind-Blowing
A blind man's visual cortex learned to "talk" to an AI that designs his artificial sight; DeepMind's AlphaGenome 2 reads DNA cell by cell to explain why one mutation can devastate one tissue and leave others untouched; a tumor-seeking drug cleared aggressive pancreatic cancers in mice and taught the immune system to remember them; MIT engineers compressed weeks of lab bioengineering into a single injection that reprograms immune cells inside the body; and Google's humanoid robot just hit a 96% success rate on a live Mercedes-Benz production line.
🔊 Industry Insights & Updates
The theme this week is deployment. NVIDIA's GR00T 2 learns new skills from just five human demonstrations and is being progressively open-sourced; ASUS put its first humanoid on a real motherboard factory line; logistics giant XO signed for up to 1,000 Apptronik Apollo robots; and Diraq installed the first quantum computer inside a commercial data center.
🧬 BioTech
The biology this week is unusually hopeful. An ALS therapy in human trials restored communication between dying neurons, with potential reach for the 97% of patients who currently have no disease-modifying option. Salk scientists built a light switch for the cell's recycling system and caught it eating half-built nuclear pores. And Gladstone researchers discovered the blood-brain barrier's real problem in Alzheimer's is a broken delivery service — one the liver can help repair.
💡 Products/Tools of the Week
This week's toolkit is all about removing friction from creative and technical work. Agent Beacon turns every AI coding session into reusable knowledge across 20+ harnesses; afk.ai rigs 3D characters in a browser with zero manual tweaking; Guidie turns a single screen recording into polished how-to guides complete with video; and VSDX.ai converts messy documents into editable, native Visio diagrams. Small tools, serious leverage.
🎥 Video Section
And for your screen time: three deep dives worth queuing up — Arctos Robotics on why an open-source PCB could change DIY robots forever, Imagination in Action on why AI investors are split on the humanoid robot boom, and Barrett's reality check on how China's robots are already working.
Looking at this week's stories side by side, it's hard not to feel a genuine thrill about where we're heading. AI is learning the language of the brain, robots are earning their keep on real factory floors, and medicine is beginning to edit the immune system like software. The future we used to describe in the abstract is being engineered, tested and deployed right now — and it's a privilege to watch it unfold week by week. Stay hungry, stay futurish!
🤯 Mind-Blowing
AI just learned to speak the visual cortex's language — and it's teaching bionic eyes to see. For a blind person, a visual cortex implant can only create phosphenes — flashes and shapes of light — and making those dots spell out useful vision has defeated researchers for decades, because everyone's cortex responds differently to electrical current. A team from UC Santa Barbara, ETH Zurich and Miguel Hernández University attacked the problem with deep learning: at a hospital in Spain, a 27-year-old man who lost his vision to traumatic brain injury received a temporary 96-electrode array in his visual cortex, and an AI model learned to predict exactly how his neurons respond to each stimulation pattern. The result: AI-designed patterns reproduced target neural activity more accurately than conventional approaches, using less electrical current — meaning sharper, safer artificial sight. The technique could make future visual prostheses precise, predictable, and personal.
DeepMind built an AI that reads DNA cell by cell — and shows why one mutation kills one cell type but spares the rest. Your heart, brain, and immune system share the exact same DNA sequence — so how can a single-letter change cause disease in one and leave the others untouched? AlphaGenome 2, announced November 11, tackles that question directly: trained on over 100 million cell profiles spanning 68 tissues, it predicts how any DNA variant changes the behavior of specific human cell types and cell states. In benchmarks, it recovered 30% of causal single-cell genetic effects versus 16% for the previous state of the art, and it's free for non-commercial research. Disease genes don't act in a vacuum — now there's an AI that knows which cell you're asking about.
A tumor-seeking drug tears down cancer's defenses — and teaches the immune system to remember. TGF-β is one of cancer's favorite shields, but blocking it body-wide hits hundreds of healthy processes. MedPacto and Case Western Reserve researchers built FEBP-sTβRII, a drug that homes in on fibronectin — a protein abundant in tumor tissue — and releases its TGF-β blocker only inside the tumor. In mice, it touched just 25 genes versus 693 for the untargeted version, cleared aggressive pancreatic tumors in 75% of treated animals, and left something remarkable behind: cured mice later rejected the same cancer, and even bladder and lung tumors they'd never encountered.
One injection could soon rewrite your immune system to hunt cancer — no lab, no surgery, no cell extraction. Today's CAR-cell cancer therapies are marvels of bioengineering with a brutal catch: immune cells must be drawn out of a patient's blood, genetically engineered in a lab for weeks, and infused back at a cost of hundreds of thousands of dollars. MIT engineers led by Daniel Anderson and Robert Langer have compressed all of that into a single infusion. Their lipid nanoparticles — antibody-fragments coated to home in specifically on natural killer cells — deliver mRNA instructions that convert those cells into CAR-equipped cancer killers while still inside the body. In mice, the reprogrammed cells wiped out cancerous T cells in lymphoma models and, remarkably, reversed liver damage in autoimmune hepatitis by clearing the autoreactive cells driving it.
Google's robot AI just scored 96% on a live Mercedes-Benz production line. Google DeepMind announced Gemini Robotics 2.5 on November 10 — a vision-language-action model that lets robots understand what they see, reason through tasks, and act in the physical world. The proof: an Apollo 2 humanoid autonomously sorting automotive parts from bulk bins into assembly-ready trays on live Mercedes-Benz lines in Kecskemét, Hungary and Sindelfingen, Germany, hitting a 96% success rate on the kitting task. Notably, the models improved dramatically after just weeks of post-training and carried over to the factory floor with very little drop-off. One honest caveat analysts flagged: Google hasn't disclosed trial duration or a commercial deployment timetable — the last 4% is the real test.
🔊 Industry Insights & Updates
NVIDIA's GR00T 2 lets robots learn new skills just by watching humans. Unveiled November 9 at CoRL in Austin, GR00T 2 is NVIDIA's next-generation robot foundation model — and its party trick is human-to-robot skill transfer. Trained partly on 100,000 hours of egocentric human video, it achieved 95% success on skill transfer after seeing just five demonstrations, and 40% zero-shot success on 50 tasks that never appeared in the target robot's training data. A compressed "Turbo" version runs on a Unitree G1 humanoid, predicting moves in 124 milliseconds on onboard compute. Better still, NVIDIA is progressively open-sourcing the model, its data, simulation environments, and even data-collection hardware — the training pipeline for physical AI is becoming a public asset.
ASUS — the laptop maker — just put its first humanoid robot to work on a real factory line. Gallop, ASUS's first humanoid, is learning to pack the fifteen-part accessory bundle that ships with a motherboard — tiny components sealed in thin plastic bags, at a workstation built for human hands. The target: 99.5% task success in 20 seconds, and the whole robot runs on a single NVIDIA Jetson AGX Thor computer. The pitch is deliberately unglamorous: prove humanoids can consistently perform real factory operations rather than staged demos, with the roadmap pointing to 2028. When the company that supplies the world's motherboards automates its own lines, that's the signal changing from demo to deployment.
XO and Apptronik just signed for up to 1,000 humanoid robots. The world's largest pure-play contract logistics provider has signed a letter of intent for as many as 1,000 Apollo humanoid robots over four years, under a Robotics-as-a-Service model that lets warehouses scale usage without buying fleets. Initial North American deployments begin in late 2027. Apollo already completed a pilot for autonomous small-goods picking, and 50+ logistics tasks — goods-to-person, case movement, packing, sortation — have been evaluated at Apptronik's Robot Park training facility. Coming right after GXO's Digit expansion, it's a second major humanoid commitment from the same warehouse operator within weeks.
A quantum computer is now running inside a commercial data center — a world first. Sydney startup Diraq has installed the first silicon spin quantum computer inside an Equinix commercial data center, sitting alongside ordinary servers and GPUs rather than in a physics lab. The current 8-qubit system targets 256 qubits within a year, draws under 20 kW, and occupies 170 square feet — with utility-scale versions projected at ~300 kW and 50 square feet. Because silicon spin qubits use standard CMOS chip fabrication, the path to millions of qubits runs through existing fabs. Diraq is also one of six companies in the final stage of DARPA's Quantum Benchmarking Initiative, with funding potentially worth $300 million.
🧬 BioTech
An ALS therapy in human trials just proved it can restore the conversation between dying neurons. In motor neuron disease, the protein TDP-43 misbehaves and corrupts the splicing of UNC13A — one of the strongest known ALS risk genes — crippling the synapses through which neurons talk to each other. Now a King's College London, UCL and Crick Institute team has shown in Nature that antisense oligonucleotides correcting that exact splicing error restore functional UNC13A protein and rescue neuron-to-neuron communication in human cellular models. The clinical translation is already live: a related therapy, TRCN-1023, is dosing patients in Phase 1/2 trials — with a potential reach of the 97% of ALS patients who currently have no disease-modifying option.
Researchers built a light switch for the cell’s recycling system — and found it eating the cell’s doors. Salk Institute scientists created ASAP, an optogenetic tool that blocks autophagy — the cell’s recycling machinery — within minutes of a flash of light, reversibly and precisely. Screening what accumulates in those first minutes revealed surprise targets: nuclear pore complexes, the gatekeepers of the cell’s nucleus, are rapidly regulated by autophagy — and only half-built pores get eaten while healthy ones are spared. It’s also the first direct link between protein translation and autophagy. Timely note: optogenetics just won the 2026 Nobel Prize in Medicine, and this tool is headed at cancer, neurodegeneration, and aging.
The blood-brain barrier isn’t just a wall — it’s a delivery service, and Alzheimer’s breaks the delivery before anything leaks. Gladstone Institutes researchers mapped every protein the blood-brain barrier imports from blood into the brain — 92 out of 861 tracked, most manufactured in the liver. The surprise in Cell: in Alzheimer’s mice, the barrier stops importing protective proteins like APOA1 before any physical breakdown or leakage begins. When the team engineered mice to produce extra liver APOA1, Alzheimer’s signs diminished. The failing transport system is a cause, not a casualty — and the liver-to-brain supply route is a treatment target that never has to cross the barrier it’s fixing.
💡Products/tools of the week
A cross-harness memory layer for AI coding agents has been released as open source, capturing full session history across Claude Code, Cursor, Codex, OpenCode and more than 20 other harnesses. The tool, Agent Beacon, works locally first and identifies useful workflows, corrections and debugging patterns from past sessions, then makes that knowledge reusable by future agents through MCP or Agent Skills. All captured activity is normalized into a common OpenTelemetry-based event model, giving developers one searchable dataset regardless of which coding tool they use, so knowledge compounds across tools and sessions instead of disappearing when a session ends.
A browser-based AI auto-rigging tool now places skeletons and paints skin weights on uploaded 3D models without manual adjustment. The tool, afk.ai, handles characters, animals, creatures, robots and props across fourteen skeleton templates including humans, quadrupeds, dragons, birds, snakes, fish and dinosaurs, and lets users preview and select from over 170 animations before exporting rigged GLB or FBX files ready for Unity, Unreal Engine, Godot, Blender or Three.js. An npm package and MCP server are also offered for developers who want to rig models programmatically or fold auto-rigging into AI workflows.
A single recording of a task is all it takes to produce a full how-to guide, thanks to Guidie, a browser extension that outputs written step-by-step instructions, annotated screenshots and a narrated MP4 video in one pass. Once the user presses Start in the Chrome extension and performs the task, every click, field and choice becomes a step with its own screenshot, and Claude writes one plain-language instruction per step along with a title, summary and tips when the recording stops. Typed text is never captured, form fields are masked before screenshots, and sensitive details like emails and API keys are flagged for blurring, and the resulting guides can be shared with teammates, downloaded as video, or marked with freshness tags showing when a re-recording is due.
Editable Visio diagrams are generated from spreadsheets, PDFs, Word documents, images and audio recordings by VSDX.ai, an AI-powered workspace covering nine diagram types such as flowcharts, swimlanes, architecture diagrams, BPMN, UML and ER diagrams. The tool reconstructs real shapes and connectors rather than a flat AI image, so results can be revised in an embedded draw.io editor with version history and then exported as a native .vsdx file that opens in Microsoft Visio. Uploaded sources remain private within each project, and existing VSDX files can be imported back into an editable project for further work.
