DEV Community: pablo padlo

Pollen Robotics Open-Sources a 25cm, 800g Reinforcement-Learning Robot Duck

pablo padlo — Tue, 20 Oct 2026 09:16:35 +0000

Pollen Robotics Open-Sources a 25cm, 800g Reinforcement-Learning Robot Duck

Pollen Robotics, the Hugging Face-backed open-source robotics company, has released Microduck: a bipedal robot just 25cm tall and 800g in weight whose walking, recovery, and object-manipulation skills are trained largely through reinforcement learning rather than hand-coded motion. Both the robot's runtime and its RL training pipeline are now open source, making Microduck one of the smallest fully open embodied-AI development platforms on the market — and a compact reference implementation for an entire stack: MuJoCo simulation, PPO training, Sim2Real transfer, ONNX export, and real-time control in Rust.

Watch the demo (video):

https://x.com/i/status/2093175236367905091

What 25cm and 800g Can Actually Do

The numbers in the Microduck release describe a robot that punches well above its size. The hardware runs on a Rockchip RK3566 — an inexpensive ARM system-on-chip more commonly found in set-top boxes than robots — driving 15 servos at a 50Hz control frequency. That control rate matches the frequency at which the RL policies are trained, meaning the policy that emerges from simulation is deployed essentially unchanged.

The skill set, per the release details, goes well beyond basic locomotion:

Capability	How it works
Walking and gamepad steering	Core RL locomotion policy, 50Hz on real hardware
Recovery from falls	Separate trained skill (get-up behavior)
Object pickup	Bend-down-and-grasp skill
Sit, stand, kick, forward roll	Discrete trained actions
Wheeled/gliding mode	Second RL policy, engaged when skate attachments are fitted
Downhill descent and spinning on skates	Same wheeled policy, distinct from the walking policy

The wheeled variant is a telling detail: attaching skates to the feet is detected, and the robot switches to an entirely separate RL policy for gliding, spinning, descending slopes, and rising while on skates. This is not a single model stretched to cover edge cases — it is modular policy composition, the way production robot fleets are actually built.

A Training Pipeline Built for Sim2Real

The open-sourced training project, microduck_rl, follows a pipeline that has become the industry standard for learned locomotion: MuJoCo and MuJoCo Warp for simulation, PPO as the reinforcement-learning algorithm, domain randomization for robustness, then Sim2Real transfer via ONNX export to the physical robot.

The noteworthy engineering is in the fidelity of the simulation. To close the sim-to-real gap, the training environment models phenomena that cheaper robotics projects routinely ignore:

Motor friction
Battery voltage variation
Control latency
Servo gear backlash

Gear backlash in particular — the mechanical slack between meshing gears in cheap servos — is a classic source of failed transfers, because a policy trained in an idealized simulator learns control strategies that depend on precision the hardware cannot deliver. Simulating it explicitly is the kind of unglamorous work that separates demos from deployable systems.

The Software Architecture Is the Quiet Story

The robot-side software is written mostly in Rust and deliberately structured as small, single-purpose services rather than a monolithic framework. The disclosed services include robotd (50Hz real-time control, RL policy, safety), updaterd (OTA updates with signature verification and automatic rollback), configd (Wi-Fi and device identity), btd (Bluetooth), padd (gamepad), mediad (camera, audio, WebRTC), and tofd (time-of-flight depth sensing).

Two design decisions stand out. First, services communicate over Unix Sockets with JSON-RPC 2.0, and only robotd can command motors. Phones, gamepads, and custom programs transmit only "action intents," which pass through robotd's safety checks before execution. This is a safety-oriented permission model — the agent layer can request, but only the certified control layer can act — an architecture that larger humanoid programs would do well to emulate.

Second, the OTA system implements the full enterprise update flow: download release, verify signature, switch version, restart, health check, and automatic rollback if the health check fails. A failed update bricks neither the software nor the robot.

Why It Matters

Microduck sits at the intersection of three trends reshaping embodied AI.

Cost curves. The entire compute stack — an RK3566-class SoC, 15 servos, a 50Hz control loop — is built from commodity components with no exotic hardware. When a reinforcement-learned robot can run on parts priced for consumer electronics, the floor for experimentation drops sharply, and that floor is what determines whether robotics research remains concentrated in well-funded labs or disperses to hobbyists, students, and small teams.

Open source as a distribution strategy. Hugging Face's backing of Pollen Robotics signals a bet that the path to useful robots runs through open models and open tooling, analogous to how open-weight LLMs compressed the time between research publication and production deployment. Releasing both the runtime and the training code means a researcher can reproduce the full loop — train in MuJoCo, export to ONNX, deploy on hardware — without reverse-engineering anything.

A reference platform, not just a toy. The combination of RL training, Sim2Real discipline, ONNX deployment, real-time Rust control, WebRTC media streaming, and robust OTA in a 25cm package gives Microduck outsized instructional value. It is, in effect, the architecture of a serious robot compressed into a desktop form factor. For anyone studying reinforcement learning, robot control, or embodied intelligence, the design document is the device.

The official release notes close with a disclaimer: no ducks were harmed in the making of the robot, though several were consulted. The joke lands because the underlying project is earnest — a small, cheap, thoroughly engineered platform that makes learned robot behavior something a developer can download, train, and watch stumble, recover, and roll forward on their own desk.

Sources

Original post by @YM_AlphaNotes: https://x.com/i/status/2093175236367905091

Join GPT Brunch — AI & business community in Cyprus: https://t.me/+H1S-YWnEIxhiMmE6

A 90-Year-Old Veteran Walks in an Exoskeleton and Reports Feeling Years Younger

pablo padlo — Sat, 17 Oct 2026 20:59:53 +0000

A 90-Year-Old Veteran Walks in an Exoskeleton and Reports Feeling Years Younger

Watch the demo (video):

https://x.com/i/status/2090638844152717373
Watch the demo (video):

https://x.com/i/status/2090638844152717373

What the demo actually shows

The significance of the clip is less about raw capability and more about fit, tolerance, and user experience. Exoskeletons have historically struggled with a trade-off: adding mechanical power to a human body requires rigid frames, motors, batteries, and control systems, all of which add weight and pressure points. A younger, trained user can adapt; a 90-year-old cannot. That the veteran described the device as completely comfortable — and subjectively felt younger — suggests the hardware achieved three things at once: low perceived weight, unobtrusive actuation, and intuitive wearability.

A few observations stand out in the demonstration footage:

Donning simplicity. An elderly user was able to put the device on and operate it, implying a design that does not require technician assistance or calibration routines.
Comfort at the interface. Skin, joints, and old bones are unforgiving test surfaces. Complete comfort implies padded, well-distributed contact points rather than rigid strapping.
Psychological effect. Reporting that he felt many years younger is not a marketing line — it indicates the device delivered genuine mechanical assistance rather than merely not hindering movement. Assistive technology that restores a sense of agency can improve willingness to walk, which in turn supports the mobility feedback loop clinicians care about.

Quantifying the test

The post provides a small set of figures, but they are striking in combination:

Attribute	Value reported in the demo
Age of the wearer	90 years
Comfort rating	"Completely comfortable"
Perceived effect	"Many years younger"
User type	Untrained civilian, not a lab subject
Setup assistance implied	None reported

The most important number is the first one. Validating wearable robotics on a 90-year-old is a far stricter acceptance test than validating on a 30-year-old engineer. Comfort complaints that a young user filters out — pressure, heat, noise, the sense of being "worn by the machine" — surface immediately in older users. The fact that none surfaced here is the strongest data point in the clip, even though it is qualitative rather than measured.

It is also worth noting what the demo does not quantify: walking speed, battery life, load-bearing capacity, and price are all absent. Those omissions define where skepticism should remain.

Why it matters

Cost curves. Exoskeleton economics have long been hostage to low production volumes. Devices built for rehabilitation clinics or factory floors sell in small numbers at high unit costs. A wearable comfortable enough for a 90-year-old points toward a mass-market assistive category — aging adults who want to walk unaided for longer. Consumer scale, if achieved, pulls down component costs (actuators, batteries, lightweight alloys, force sensors) the same way any high-volume electronics category does. The veteran's comfort report is therefore not just a product anecdote; it is evidence that the form factor may finally fit a market large enough to bend the cost curve.

Adoption. The single biggest failure mode for assistive devices is abandonment: users acquire them, find them stigmatizing or uncomfortable, and leave them in a closet. "Completely comfortable" plus "felt many years younger" attacks both objections at once — physical tolerance and emotional acceptance. A device an elderly person is happy to wear is a device that gets used daily, and daily use is where clinical and commercial value concentrates.

Competition. The exoskeleton field spans industrial lifting suits, military load-carriage systems, and medical rehabilitation rigs. The demo suggests a fourth track — consumer mobility augmentation — may be opening, and it plays to manufacturing strengths in cost reduction and rapid iteration. Competitors in Japan, Europe, and the United States with clinical-grade products may find the low end contested by devices optimized not for peak torque but for all-day wearability. That is a different engineering target, and arguably a harder one.

Open questions. Comfort in a short demo is not the same as comfort across a full day, and feeling younger is not the same as measured fall reduction or gait improvement. Battery endurance, weight, durability over months of use, and regulatory classification for medical claims all remain unaddressed by the footage. Buyers and investors should treat the clip as a compelling usability proof point, not a complete validation.

Sources

Repost with the native video demo: https://x.com/i/status/2090638844152717373
Original post by @XinhuaChinese: https://x.com/i/status/2090638844152717373

Join GPT Brunch — AI & business community in Cyprus: https://t.me/+H1S-YWnEIxhiMmE6

Chinese Scenic Areas Deploy 2-kg Exoskeleton Walkers That Deliver 10 kg of Assistive Force

pablo padlo — Sat, 17 Oct 2026 20:52:27 +0000

Chinese Scenic Areas Deploy 2-kg Exoskeleton Walkers That Deliver 10 kg of Assistive Force

Multiple tourist destinations across China have begun renting lightweight mechanical exoskeleton walkers weighing just 2 kilograms that provide up to 10 kilograms of climbing assistance, marking the first large-scale commercial deployment of such devices outside industrial and military settings. The walkers, which strap onto a user’s legs and lower back, use electric motors to amplify the wearer’s own muscle power during steep ascents, effectively turning a strenuous mountain hike into a manageable stroll. The rollout during the National Day holiday period signals that exoskeleton technology has crossed a critical threshold from laboratory prototype to consumer-grade product, with implications for the broader mobility-assistance market.

Watch the demo (video):

https://x.com/i/status/2080801515904876710

Quantifying the Deployment

The exoskeleton walkers now in service at scenic areas such as Qiyun Mountain in Anhui Province, Mount Emei in Sichuan, and the Mutianyu section of the Great Wall represent a specific class of assistive device. According to reports from Chinese state media, the units weigh approximately 4 jin (2 kg) and deliver about 10 kg of assistive force, a ratio of 5:1 in output-to-weight. The table below compares the key specifications of these tourist-oriented exoskeletons with typical industrial exoskeletons used in manufacturing and logistics.

Metric	Tourist Exoskeleton (Qiyun Mountain / Mount Emei)	Typical Industrial Exoskeleton (e.g., EksoVest, SuitX)
Device weight	2 kg	4–6 kg
Assistive force	10 kg (per leg)	5–15 kg (per arm or back)
Battery life	~4–6 hours (estimated from rental duration)	8–12 hours
Rental cost	Not publicly disclosed	N/A (sold to enterprises)
Primary use case	Mountain climbing, stair assistance	Factory assembly, warehouse lifting

The 2-kg weight is particularly notable because it removes the primary barrier to consumer adoption: previous exoskeleton prototypes often weighed 8–15 kg, making them impractical for casual tourism. The Chinese devices achieve this reduction through carbon-fiber frames and compact brushless motors, a design choice that mirrors the miniaturization trend seen in consumer drones and power tools.

Why It Matters

The commercial rollout of exoskeleton walkers in tourist attractions has three major implications for the technology’s cost curve, adoption trajectory, and competitive landscape.

Cost curve compression. Industrial exoskeletons typically sell for $5,000–$20,000 per unit, pricing them out of the consumer market. The Chinese tourist-exoskeleton program, by contrast, operates on a rental model—visitors pay a per-hour or per-day fee rather than purchasing the device. This shifts the economics from capital expenditure to operational expenditure, allowing the manufacturer to amortize hardware costs over thousands of rentals. If the rental price is, say, ¥50–100 per hour (roughly $7–14), the break-even point for a ¥10,000 device would be 100–200 rental hours, which a single unit could achieve in a few weeks during peak season. Such a model could accelerate mass production and drive unit costs below ¥5,000 within two years, making exoskeletons affordable for individual purchase.

Adoption beyond tourism. The same lightweight exoskeleton technology that helps tourists climb mountains can be adapted for elderly mobility assistance, rehabilitation therapy, and last-mile delivery. China’s aging population—over 300 million people aged 60 or older by 2035—represents a massive addressable market for devices that reduce fall risk and extend independent living. The tourist deployments serve as real-world validation of reliability, battery life, and user acceptance, de-risking investment in these adjacent markets. Competitors such as Japan’s Cyberdyne and the U.S.-based Ekso Bionics have focused on medical and industrial niches; the Chinese approach of starting with tourism could give local manufacturers a lead in consumer-grade exoskeletons.

Competitive pressure on global players. Chinese manufacturers have a history of scaling production rapidly once a technology proves viable—drones, electric scooters, and solar panels are precedent cases. If exoskeleton walkers follow a similar trajectory, global incumbents may face price erosion and market-share loss. The fact that multiple scenic areas—Qiyun Mountain, Mount Emei, and Mutianyu Great Wall—simultaneously adopted the devices suggests a coordinated supply chain, likely backed by a single manufacturer or a consortium. The Xinhua article notes that the technology “grabbed the spotlight” during the Golden Week holiday, indicating strong government and media support for the initiative.

The immediate takeaway for business readers is that exoskeletons are no longer a distant future technology. They are being rented by tourists today, at a weight and price point that makes them viable for mass-market use. The next 12–18 months will determine whether this becomes a niche gimmick or the beginning of a new consumer hardware category.

Sources

Cutting-edge tech powers smart National Day travel in China: http://en.people.cn/n3/2026/1010/c90000-20506847.html
China's Qiyun Mountain scenic area offers exoskeleton robot rentals to deliver smarter mountain climbing experience: https://www.globaltimes.cn/page/202610/1371699.shtml
Advanced technologies deliver dazzling experiences at tourist destinations across China during National Day holiday: https://en.people.cn/n3/2026/1008/c90000-20505997.html
约10公斤助力！这个国庆，4斤重“登山外挂”火到峨眉山、慕田峪长城等知名景区！外国游客也来尝鲜: https://www.mrjjxw.com/articles/2026-10-07/4597855.html
租外骨骼设备爬长城？前沿科技“抢镜”黄金周说明了什么: https://app.xinhuanet.com/news/article.html?articleId=202610079fdb0594ddc945a190e1f2d286e42c0a
Original post by @luobin666888: https://x.com/i/status/2080801515904876710

Join GPT Brunch — AI & business community in Cyprus: https://t.me/+H1S-YWnEIxhiMmE6

XPeng's Humanoid Robot Walked So Naturally That Viewers Swore It Was a Person in a Suit

pablo padlo — Tue, 13 Oct 2026 11:37:50 +0000

XPeng's Humanoid Robot Walked So Naturally That Viewers Swore It Was a Person in a Suit

A video of XPeng's humanoid robot walking without its white fabric covering has passed 488,000 views on X, and the reason is uncomfortable for the robotics industry: the machine's gait looked too human. After the Chinese EV maker unveiled its next-generation Iron robot at the 2025 XPeng AI Day on November 5, more than half of the online comments questioned whether a real person was hidden inside the demonstration shell, chairman and CEO He Xiaopeng said the next day. His answer was to publish raw, unedited footage of the bare mechanical chassis walking — the clip now circulating as proof.

Watch the demo (video):

https://x.com/pgol80/status/2109970750019170327

Watch the demo (video):

https://x.com/IamRamenPanda/status/1986289235700424997

What XPeng Actually Showed

The second-generation Iron is a departure from the boxy first prototype unveiled at the 2024 AI Day, a 178 cm, 70 kg machine built around XPeng's Turing AI chip. The new unit has a humanoid spine, bionic muscles, flexible full-coverage skin, and a distinctly female figure; at the event it walked a catwalk with what witnesses described as natural, fluid movement. XPeng's R&D team attributes the light, gentle stride to passive degrees of freedom added at the toes, letting the foot roll the way a human foot does instead of slapping the ground flat.

Under the skin, Iron runs on three in-house Turing AI chips delivering 2,250 TOPS and a Vision-Language-Action model that handles seeing, moving, and conversing in real time — the same neural-network approach XPeng uses in its cars. Each hand has more than 20 degrees of freedom.

The skepticism was not confined to comment sections. XPeng escalated its proof: He Xiaopeng shared an unedited Weibo video pointing out the wind noise from internal cooling fans and the microphone array near the robot's ears, and at a follow-up event engineers cut the robot's leg open on stage to expose the mechanical components. Even Elon Musk weighed in, telling Chinese media that "Tesla and China companies will dominate the market."

Why the Doubt Is the Story

Accusations of teleoperation or human-in-a-suit follow every impressive humanoid demo — Tesla's Optimus and 1X's Neo faced the same charges. What changed here is the perceptual threshold: when a robot walks so well that the public defaults to "there must be a person inside," bipedal locomotion has crossed from laboratory curiosity into plausible product. For procurement teams and investors, that threshold matters more than any spec sheet, because it signals the technology is approaching reliability in unstructured environments.

The business stakes are large. XPeng believes the robot market could exceed the car market, and it is moving from demonstration to manufacturing: in September 2026 the company opened Iron's production line, with the robot autonomously walking off the line after assembly. Mass production is planned by the end of 2026, with official launch and deliveries in China and overseas markets scheduled for 2027.

What to Watch Next

A catwalk is a controlled stage, not a warehouse. The tests that now matter are endurance under load, navigation on uneven terrain, and hand dexterity on real tasks — plus the price tag, which XPeng has not disclosed. But the viral skepticism itself is a leading indicator: the industry has spent years making robots look impressive, and XPeng's problem is that it finally made one look ordinary.

Sources

Join GPT Brunch — AI & business community in Cyprus: https://t.me/+H1S-YWnEIxhiMmE6

Humanoid Robots Are Leaving the Stage and Clocking In on the Factory Floor

pablo padlo — Sat, 10 Oct 2026 21:01:44 +0000

Humanoid Robots Are Leaving the Stage and Clocking In on the Factory Floor

A wave of recent footage shows the same pattern across three kinds of machine: a biped walking a lab floor, a quadruped picking its way through an industrial hall, and a robotic arm working a line. The clips are collected in a single post whose framing is the honest one — the humanoid race "is about more than looking human." The unglamorous part, and the part that decides whether any of this ships, is balance, grip and control sustained over an entire shift rather than for the fifteen seconds of a demo.

Watch the demo (video):

https://x.com/i/status/2109007187318391246

What the clips actually show

The montage is not one product launch; it is a cross-section of where the category stands. A white humanoid walks unprompted across a workshop floor. A second machine performs on a dark stage, holding a pose for an audience. In an industrial hall, a wheeled quadruped — the same body plan popularised by inspection robots — moves under fixed equipment, and a tracked arm positions itself to work. Each clip is short, and each shows a machine that has solved locomotion in a controlled setting. None of them shows a full working day.

That gap is the story. Walking on two legs already took the field decades of falling over; the harder problem is doing it not once, on cue, but repeatedly, under load, when the floor is uneven and nothing is scripted. The source post compresses this into a single line — the demo is easy, a real shift is not — and the footage supports it: every clip is a capability proof, not an endurance run.

The demo-versus-shift gap

It helps to separate what a robot can do from what it can do repeatably. The difference is not cosmetic; it is the difference between a research budget and a purchase order.

Dimension	The demo	A real shift
Locomotion	scripted, flat ground	stairs, ramps, uneven floors
Payload	none or symbolic	continuous material handling
Balance	reset after each take	maintained for hours
Control	pre-programmed sequence	real-time correction under disturbance
Environment	stage or lab	live floor with people and machines

Read down the right-hand column and the bottleneck stops looking like software. Balance under a changing load, grip that survives dust and wear, and actuators that hold tolerance after a thousand cycles are hardware problems — motors, gearboxes, sensors and thermal budgets. This is why the industry's announcements increasingly read like engineering updates rather than AI updates, and why independent coverage keeps landing on the same conclusion: AI breakthroughs in robotics won't change your life any time soon until the reliability work is done.

Why it matters

For buyers, the practical question is shifting from "can a robot do this task?" to "can it do it every shift, for a known cost?" That reframing favours machines that integrate into existing layouts without rebuilding them: if a platform can climb the stairs already in a plant, it avoids the infrastructure bill that has sunk earlier automation projects. Coverage of how humanoid robots navigate workplaces makes the same point from the facilities side — the building, not the brain, is often the constraint.

For the wider field, the pressure is competitive. Boston Dynamics has spent years moving from spectacle to work, with AI-powered robots learning factory tasks, while newer entrants such as Figure have pushed autonomous ladder climbing as a milestone that matters to customers, not judges. Against that backdrop, a montage of walking humanoids, a quadruped and an arm is best read as a status report: the choreography is solved, and the new job — the one that pays — is only beginning.

Sources

A Humanoid Walked Into a Furnace: What Figure's F.02 Send-Off Really Says

pablo padlo — Thu, 08 Oct 2026 12:33:42 +0000

A Humanoid Walked Into a Furnace: What Figure's F.02 Send-Off Really Says

Most robots end their working life in a storage room or a skip. This one ended it in a 75-ton electric arc furnace in Finland, on purpose, after being trained to jump into it.

Figure, the humanoid robotics company, retired most of its F.02 fleet this month — and doing it properly turned into an engineering project of its own. The full story is on Figure's own site, and it is stranger than the clip suggests.

https://x.com/pgol80/status/2105956486627180705

Why a company would melt its own robots

The F.02 was not a demo unit. It was Figure's first robot deployed at BMW, the first to run the company's Helix AI, the first to do household chores, and the first to work a logistics shift. As the newer F.03 fleet came online, keeping the older machines running stopped making sense.

The problem was what was inside them: custom-built actuators and other proprietary hardware. Disassembling each robot by hand would have tied up the technical team and pushed back the launch of the next generation, the F.04. Destroying the fleet in bulk solved both problems at once — as long as the destruction also destroyed the intellectual property.

The send-off came from the internet

Figure asked publicly what it should do with the robots. Arnold Schwarzenegger replied with the obvious answer: melt them. It is hard to imagine a better fit — his T-800 character ends the same way at the end of Terminator 2, sinking into molten steel. Figure reached out, he agreed to take part, and the idea became a plan.

Finding a foundry willing to drop a humanoid with lithium-ion batteries into expensive equipment was the hard part. Facilities in the United States and Mexico refused. Figure even contacted former MythBusters for help. The only site willing to take it on was a foundry in Imatra, Finland.

The robot had to learn to die well

This is the part that makes it more than a stunt. The F.02 could not simply be pushed into the furnace — it had to walk and jump into the furnace on its own, accurately, at a plant the team had never visited.

Training began at Figure's campus in San Jose. Engineers set up large airbags and taught the robots to jump off the second floor of a building, using the movements of stunt performers as reference. A new AI model was trained in simulation so the machines could land precisely in a vat of molten steel.

At the foundry, three graphite electrodes powered the arc furnace. The team had 24 hours and six melts, each offering only about a 20-minute window before the steel cooled into a hard crust. According to Figure, the robots kept running their control policies through the extreme heat and electromagnetic fields — even as cameras and other electronics failed around them.

What is left, and why it matters

The molten metal was shipped back to the United States, and Figure is now machining it into a limited series of commemorative artifacts, sold through its online shop. Most of the fleet is gone; a few units remain in storage at headquarters.

Two things about the story are worth keeping.

First, it is an unusually honest piece of hardware marketing. In an age where robots are routinely faked with AI-generated video, Figure went out of its way to state that the footage is real — that the machines really were trained to jump, really were shipped to Finland, and really did drop into the steel. Several outlets covered it the same way, from The AI Insider to Mezha, and the distinction is the whole point.

Second, it is a rare look at the other half of a robot's life cycle. We spend enormous attention on what a humanoid can do when it works. Almost none on what happens when it is obsolete: how you retire it, how you protect what is inside it, and how you turn an end into a story.

Figure pulled that off. It trained a machine to walk into the fire, and it made people watch.

A Spider Robot Designed by AI: Code Is the New CAD

pablo padlo — Wed, 07 Oct 2026 12:52:56 +0000

A Spider Robot Designed by AI: Code Is the New CAD

Somebody just designed a full spider robot without opening a single CAD program. No sketches, no solids, no feature tree. Every part was written as a line of code, and the whole machine fell out of a text description.

The project, shared by @MirrortekUK on X, used Claude Opus end to end. The result is a six-legged robot with two additional arms, built from five hundred parts and thirty joints. Each part is defined by a short Python program, and all of the parameters live together in a single YAML file.

https://x.com/MirrortekUK/status/2107790410969976859

What was actually built

The interesting part is not the robot itself — hexapods are old news. It is the pipeline. Because every component is code, the same description can be exported into formats that the rest of the engineering world already uses: STEP for geometry exchange, URDF for robotics, and a MuJoCo scene for physics simulation.

That is the quiet shift. A description is not a picture of a robot; it is a machine that produces the picture, the simulation, and the manufacturing files at the same time. Change a parameter — a link length, a joint limit — and the geometry, the kinematic tree, and the simulated dynamics all update together, because they were never separate things in the first place.

Code-defined hardware is a real movement

This is not a first. For years, tools like CadQuery and build123d have let engineers script solid models in Python instead of clicking through a GUI. Open-source robotics stacks have long paired a URDF description with a MuJoCo model so that one source of truth drives both control and simulation.

What is new is who writes the code. Until recently, code-defined CAD was a discipline you had to learn — a specialist skill on top of mechanical engineering. Now the model can draft the geometry program from a plain description, and the engineer reviews and refines it. The bottleneck moves from "can I model this?" to "is this the right design?"

Why the export formats matter more than the render

A pretty image proves nothing. What makes a code-defined robot real is that its output is not an image but a set of files with meaning:

STEP is the neutral exchange format every machining and manufacturing shop can read. If your design can emit STEP, it can leave the design tool and enter a supply chain.

URDF is how a robot describes its own body to its software — links, joints, limits, the kinematic tree. It is the bridge between the mechanical design and the control stack.

MuJoCo is a physics engine. A scene exported from the same source lets the designers test how the machine behaves before any hardware exists — how it stands, walks, and recovers.

When one description produces all three, you have a closed loop: describe, simulate, refine, and only then cut metal.

The honest caveats

It would be easy to overclaim here, so let us not. A model that drafts geometry still produces geometry that has to be checked: tolerances, materials, manufacturability, and load paths are engineering judgments a language model does not own. The claim is not that AI replaces the mechanical engineer. It is that the engineer's starting point is now a working draft instead of a blank canvas.

It is also worth noting that this is a single project, shared publicly and not yet independently verified. Treat the headline as a direction of travel, not a finished product.

Why this matters beyond one spider

Hardware has always lagged software in iteration speed, because hardware lives in drawings. Drawings are slow to change, hard to diff, and impossible to test without building. Code is none of those things. The moment a machine is a program, it inherits everything software is good at: version control, review, parameter sweeps, automated tests, and instant simulation.

That is why a small open-source spider robot is a better signal than a dozen flashy humanoid demos. It points at a different future: not robots that look impressive, but robots that are cheap to describe, cheap to change, and verifiable before they exist.

Code is the new CAD.

The Robot That Splits in Half: What D1's Modular Body Actually Changes

pablo padlo — Mon, 28 Sep 2026 12:24:17 +0000

The Robot That Splits in Half: What D1's Modular Body Actually Changes

A robot that is two robots, or one, depending on a coupler. D1 — built by Hong Kong's Direct Drive Technology — ships as two 24.3 kg wheeled-legged modules joined by a magnetic coupler at the centre of the body. Docked, they form a 48.5 kg quadruped. Split, they are two independent bipeds, each carrying its own battery and its own computer. Nobody has to touch them to make the switch.

That is what the footage tests, and it holds: the machine re-docks itself with no human hand involved and keeps moving.

The body is the configuration

Most legged robots answer a single terrain. Bipeds are efficient on flat ground; quadrupeds are stable on the rough stuff. D1 declines to choose. A single module is a wheeled biped rated at 2 m/s. Clamp two together and the pair becomes a 48.5 kg quadruped rated at 3.2 m/s — roughly 11 km/h — with up to 100 kg of payload prone and 80 kg standing. The same hardware covers both cases; the difference is a coupling decision.

Modularity here is not a marketing word for "upgradeable". The legs, wheels, sensors and arms are all designed to attach to the same core, and a second module is a peer rather than an accessory: it brings its own compute, its own battery and its own autonomy, which is why the split state is not one robot limping in two pieces but two working machines.

What's inside each module

Per module: an NVIDIA Jetson Orin NX 8 GB running Ubuntu 22.04 with ROS 2 support — the same stack research teams already use. A 43.2 V / 9 Ah pack (388.8 Wh) gives at least five hours of no-load runtime and up to 25 km of travel, recharged in under two hours. Rolling speed tops out at 11 km/h in the docked configuration; the manufacturer lists 2 m/s for bipedal mode and 3.2 m/s for quadrupedal. The docked pair clears obstacles up to 70 cm, with 50 cm recommended as the working limit.

Direct Drive Technology also publishes its own ROS 2 SDK, IsaacLab reinforcement-learning environments and a sim2real bridge — the software side is deliberately open, which tends to matter more for adoption than any single spec.

Why modularity is the real claim

The interesting number is not the payload; it is what the reconfiguration removes. Instead of buying a biped for corridors and a quadruped for rubble, an operator buys one platform and lets it change shape to suit the ground in front of it. Fewer machines, fewer spares, one software stack — the argument is maintenance and inventory, not spectacle.

It also changes what "fleet" can mean. If modules are peers, a field team can start with a quadruped and end the week with two bipeds, with no new hardware shipped.

What to watch

"World's first fully modular robot with embodied intelligence" is the vendor's phrase, not a measured benchmark — filed next to every other robotics first-claim. Pricing, though, is public: $7,499 for a single biped, $13,999 for the pair.

The next thing to watch is not the docking trick. It is whether a group of these can re-configure mid-task, without downtime and without a human hand.

Watch the demo (video):

https://x.com/pgol80/status/2104546645732147387

Sources:

Original post by @YM_AlphaNotes: https://x.com/YM_AlphaNotes/status/2100503951951212859
New Atlas on the D1: https://newatlas.com/robotics/d1-modular-quadruped-biped-robot
Direct Drive specs and pricing: https://robotlar.org/en/dog-robots/directdrive-d1
Interesting Engineering: https://interestingengineering.com/ai-robotics/world-first-convertible-robot
Direct Drive Robotics on GitHub: https://github.com/DDTRobot

The AI That Refuses to Write: What Jev Changes When It Returns Decisions

pablo padlo — Wed, 23 Sep 2026 14:01:15 +0000

The AI That Refuses to Write: What Jev Changes When It Returns Decisions

An AI model that cannot write a sentence. Jev, released by San Francisco's TypeSafe AI on 15 September 2026, does not generate text at all. You hand it unstructured state — a support ticket, a log line, a page of messy input — plus a set of typed questions, and it returns typed answers with calibrated probabilities. No string, no parse step, no option you did not declare.

The trade is the point. TypeSafe calls the class "System One models", borrowing Daniel Kahneman's split between fast intuition and slow reasoning, and argues that RLHF — the method behind chat LLMs — optimised for human preference rather than for decisions software can trust without a human in the loop.

A model with no output string

Everything about the interface is typed. The possible answers are declared in advance, so the model cannot return a malformed value or invent an option outside the set. All answers come back in a single parallel pass rather than one token at a time, which is where the speed comes from. TypeSafe describes the result as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out".

The system has no free-form output surface at all. It cannot write a paragraph, a code block or a refusal — a limitation the company frames as the source of its advantages.

What the numbers say — and who measured them

The benchmarks are TypeSafe's own, and the coverage repeats them as such. End-to-end latency is listed at 70–500 ms, against a 3–329 s range for frontier LLMs on the same System One shaped tasks. Input pricing is $0.042 per million tokens; output is free, described as too cheap to meter. Training uses RLCD — Reinforcement Learning for Calibrated Decisions — on a new architecture the company has not disclosed.

Independent write-ups, from Tom's Hardware to DataCamp, carry the same caveat: these are vendor figures, not third-party measurements.

Why calibration is the real claim

"Can't hallucinate" is the headline, and it deserves a careful read. It is a schema guarantee: Jev cannot produce a malformed value or an option outside the declared set. It can still pick the wrong valid option, and TypeSafe's own FAQ acknowledges as much. What is meant to surface those cases is calibration — the property that a model saying 90% is right about 90% of the time. A model that says 90% and is right 60% of the time cannot be automated around, no matter how fast it is. Calibration is measurable, and it is the claim worth auditing.

The other half of the argument is scope. TypeSafe does not pitch Jev as a replacement for general LLMs; it pitches it at the places where software currently shells out to one and then has to parse and validate the result: classify, route, score, extract, branch, guardrail.

What to watch

The demo that spread is the least serious use case. On 17 September, @_MaxBlade posted Jev playing Subway Surfers at superhuman speed — 50 games at once, under a cent for the run, every lane choice labelled with its probability. It is a clean visualisation of the primitive: a decision function, called thousands of times a second, returning numbers you can act on immediately.

The serious version of that capability is boring on camera — routing a ticket, scoring a lead, guarding another model's output. The next number that matters is not the speed multiple. It is how often Jev's confidence is right when the decision actually costs something.

Watch the demo (video):

https://x.com/pgol80/status/2102759773363978682

Sources:

Original demo by @_MaxBlade: https://x.com/_MaxBlade/status/2100634359099232678
TypeSafe AI launch post: https://typesafe.ai/blog/introducing-system-one-models-and-jev
Tom's Hardware: https://www.tomshardware.com/tech-industry/artificial-intelligence/typesafe-ais-jev-offers-an-alternative-to-llms-that-claims-to-be-193x-faster-and-445x-cheaper-system-one-type-model-is-bespoke-for-probabilistic-decision-making
MarkTechPost: https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/
DataCamp explainer: https://www.datacamp.com/blog/system-one-models-jev

What an Hour of NVIDIA's B300 Actually Costs in 2026

pablo padlo — Tue, 22 Sep 2026 11:52:27 +0000

What an Hour of NVIDIA's B300 Actually Costs in 2026

Renting NVIDIA's Blackwell Ultra B300 — the GPU the 2026 AI buildout is priced against — costs between $6.94 and $8.71 per GPU-hour on demand. Spot capacity trades at $3.75, and the on-demand price has climbed about 83% in six months. This is what a single hour of the compute behind every frontier model actually costs.

The spread between providers is the first surprise

The same chip sells for three different prices depending on who is renting out the hour. RunPod's community tier lists a single B300 at $6.94 per hour; its secure tier at $7.89. Verda in Finland lists $7.50 with spot capacity at $3.75. Nebius sits at $7.85, with spot at $4.30. On eight-GPU nodes the arithmetic improves: Excesssupply advertises $36 an hour for eight B300s, which is $4.50 per GPU.

Third-party trackers give the medians buyers actually benchmark against. Cloud-gpus.com (11 September 2026) reports $8.71 on demand, $3.75 on spot and $7.24 on reserved capacity across five providers. Gambit.zone's 10 August snapshot puts the B300 median at $7.94 across seven quotes, with a range from $7.17 to $17.80. Gputable.dev tracks a cheapest-listed floor of $6.60.

The card is the cheap part

A single B300 is estimated at roughly $50,000. A DGX B300 with eight of them lists near $350,000. A full GB300 NVL72 rack — 72 B300s, 36 Grace CPUs, 37 terabytes of fast memory — is estimated between $3M and $6.5M depending on how the measurement is done.

That last figure is where the market is heading. Microsoft has put the first at-scale GB300 NVL72 cluster into production for OpenAI workloads, and Azure's ND GB300 v6 VMs are being sold as the standard infrastructure for multitrillion-parameter models.

Why the price is going up

The on-demand number is the visible artifact of a supply constraint. Across the pricing boards, on-demand B300 capacity went from roughly $4.5 an hour to $9.16 in about six months — a rise of 83%. Reserved contracts are currently around 2.7 times cheaper than the on-demand spike, a wider spread than at any point in the H100 era.

That gap is the honest signal. Both vendors and end users report supply constraints, and the width of the reserved discount tells you how much of the on-demand price is scarcity rather than cost.

What the silicon actually buys

Each B300 carries 288 GB of HBM3e, 8 TB/s of memory bandwidth and 18 PFLOPS of FP4 inference performance. The precision and the bandwidth are the point: long-context agentic inference is bottlenecked on exactly that combination, because a large KV cache has to stay in the fastest memory tier rather than being offloaded.

NVIDIA frames the workload as test-time scaling — the third scaling dimension after pretraining and post-training. The company's own documentation notes that reasoning at inference can demand up to 100 times the compute of one-shot generation, which is the demand curve the B300 was designed around.

What to watch

NVIDIA's headline figures are 35 times cheaper tokens and 50 times the throughput per megawatt versus Hopper, based on SemiAnalysis InferenceX benchmarks. On DeepSeek-R1, Blackwell Ultra systems reached 2.5 million tokens per second in MLPerf Inference v6.0, and NVIDIA quotes $0.24 per million tokens as an inference reference.

Those are vendor and partner numbers. The number that decides an actual bill is utilization: a node charges for the hour whether the silicon is working or waiting. That is not on any datasheet, and it is the one worth watching.

Watch the rack (video):

https://x.com/pgol80/status/2102364543091376230

Sources:

Live marketplace listings, 09–17 September 2026
Footage: Inventec full B300 rack, GTC 2025: https://www.youtube.com/watch?v=QPXbxzVw_As
cloud-gpus.com price analytics (11.09.2026): https://cloud-gpus.com/price-analytics
gambit.zone GPU rental medians (10.08.2026): https://gambit.zone/data/gpu-prices
gputable.dev cloud GPU price trends: https://gputable.dev/trends
Spheron cloud GPU provider comparison: https://spheron.network/blog/top-10-cloud-gpu-providers
NVIDIA GB300 NVL72 product page: https://www.nvidia.com/en-us/data-center/gb300-nvl72/
Microsoft Azure GB300 NVL72 cluster for OpenAI: https://azure.microsoft.com/en-us/blog/microsoft-azure-delivers-the-first-large-scale-cluster-with-nvidia-gb300-nvl72-for-openai-workloads/

Fifty Accounts, One Agent: The Dead Internet Theory, Filmed From the Operator's Chair

pablo padlo — Mon, 21 Sep 2026 09:02:44 +0000

Fifty Accounts, One Agent: The Dead Internet Theory, Filmed From the Operator's Chair

A video posted on X on 19 September 2026 shows a workstation that looks, at first glance, like a mobile developer's test bench: a single widescreen monitor tiled with dozens of emulated smartphone screens, each logged into its own account. Chat windows scroll in one pane, a feed refreshes in another, a reply thread updates in a third — all at the same time, none of it touched by a human hand.

The poster's claim is precise: a Chinese AI agent runs fifty social media accounts from this single setup, around the clock. Whether the number is exactly fifty matters less than what the screen makes visible — the infrastructure of the dead internet theory is not a metaphor. It is a workstation with a fleet of virtual phones.

What the theory actually says

Dead internet theory began as a 2021 forum essay arguing that most online activity had quietly become non-human — bots posting, bots replying, bots engaging bots, while real people dwindled into the audience of an automated theater. The Atlantic's Kaitlyn Tiffany gave it mainstream circulation that August under a verdict that aged remarkably well: wrong, but prescient.

The theory's core observation was never that the internet would go quiet. It was that the internet would stay loud and become empty at the same time — activity without actors, conversation without conversants. Scroll any large platform today and the prediction reads less like a conspiracy than like a changelog.

The measurable part stopped being controversial

In June 2026, Cloudflare reported that automated traffic had crossed over: bots now generate about 57.4% of web page requests on the networks it measures, versus 42.6% from humans. Cloudflare's CEO Matthew Prince had publicly predicted this crossover would arrive around the end of 2027. It came more than a year early.

Other measurement firms date the flip differently. Thales, in its Bad Bot reporting, argues bot traffic became the majority as early as 2023. Imperva's 2024 report already put non-human traffic near half of the total. The firms disagree about when it happened because nobody sees the whole web — they agree, increasingly, that it did.

The most striking number is not the share but the change in kind. HUMAN Security's 2026 benchmark reports that traffic from AI agents which take actions on the web — clicking links, filling out forms, completing multi-step tasks — grew roughly 7,851% year over year. Scrapers grew too, at 597%. The older generation of bots read the internet; the new generation performs on it. A wall of fifty phone accounts generating posts, replies and likes is not an edge case of this trend. It is the trend at hobbyist scale.

Why one workstation matters

Fifty accounts is small. That is exactly why the footage lands: nobody needs a data center anymore. Cloud-phone farms and agent frameworks have collapsed the cost of running a convincing persona from a salary to an API bill, and the economics of engagement rewards volume over identity everywhere they operate.

Researchers tracking engagement-driven bots have found them optimized for the metrics platforms monetize — followers, replies, watch time — on every major network. Platform-level estimates now floating around the industry put AI-generated content near half of new uploads in some media categories. Each estimate comes with caveats, but the direction has been consistent for three consecutive years.

The second-order effect matters more than the first. When bots are the majority of traffic and agents are the fastest-growing share, recommendation systems train on synthetic behavior, advertisers price synthetic attention, and the feed becomes a mirror of the machine. The dead internet was never about the absence of people. It is about the absence of evidence that anyone is there.

What to watch

The footage itself is unverifiable at this distance — one screen, one poster's claim, no look at the agent stack underneath. The honest reading splits cleanly: the screen shows real infrastructure; the autonomy claim is attribution, not demonstration. Both halves are worth tracking. Verified agentic traffic scales are public; single-operator account farms are not.

The next useful number will not be how many accounts one agent can run. It will be how long a network of them can run before the humans in the thread notice.

Sources: original post — Roundtable Space on X · NBC News on Cloudflare's bot crossover · Fortune: dead internet theory, measured · The Atlantic, 2021: the theory's ur-text coverage · The Guardian TechScape, 2024

Agentic Systems: Balancing Simplicity and Complexity in AI Engineering

pablo padlo — Mon, 21 Sep 2026 08:24:40 +0000

Agentic Systems: Balancing Simplicity and Complexity in AI Engineering

By 2026, agentic frameworks will reach a critical maturity point, yet their success often hinges on avoiding unnecessary complexity. Effective agentic systems prioritize simple, composable patterns over specialized libraries to maintain control. Data from Anthropic reveals a sharp divide: the most successful implementations distinguish between workflows with predefined code paths and agents that dynamically direct their own processes.

Defining Agentic Systems: Workflows Versus Autonomous Agents

Agentic systems generate their own search queries to direct operations, a departure from passive query-response models. Workflows follow predefined code paths, while autonomous agents dynamically determine task strategies. Workflows offer consistent latency, while agents introduce flexibility and variable execution times, increasing debugging complexity.

Augmenting an LLM with retrieval, tools, and memory enables agentic behavior without opaque frameworks. The industry now evaluates entire agentic AI systems rather than individual LLM performance, necessitating new metrics for success. Simplicity often trumps complexity, with retrieval-augmented generation frequently sufficing over full autonomy.

Constructing the Augmented LLM with the Model Context Protocol

An LLM enhanced by retrieval, tools, and memory forms the core of agentic architecture. The Model Context Protocol standardizes integrating these capabilities with third-party tools, reducing abstraction layers for clearer debugging. Tailoring augmentations to specific use cases is crucial, as generic applications risk latency spikes and inefficiencies.

Well-documented interfaces are prioritized over feature-heavy SDKs to avoid customer errors caused by incorrect assumptions. Readability beneath the interface ensures maintainability, regardless of the stack.

Predictability Versus Flexibility: Deciding Between Workflows and Agents

Fixed logic benefits from workflows, while flexible decision-making requires agents. Task volatility drives this choice, not perceived sophistication. Workflows excel with static code paths, offering consistent latency and deterministic debugging. Agents trade latency and cost for improved task performance in novel scenarios.

Feature	Workflows	Agents
Pathing	Predefined	Dynamic
Latency	Fixed	Variable
Debugging	Deterministic	Complex
Use Case	Static Tasks	Novel Scenarios

Simpler patterns serve as defaults unless task requirements demand autonomous adaptation. Frameworks like LlamaIndex offer rapid prototyping but may obscure prompt mechanics critical for production reliability.

Mechanics of Agent Orchestration and Flexible Routing

Routing logic in orchestrator-worker setups sorts incoming requests, directing them to specialized sub-processes. This avoids performance drops from generalized prompts. Flexible tool selection enables LLMs to call third-party functions based on real-time needs, moving systems from fixed code paths to dynamic loops.

Parallel execution cuts latency but increases token costs and debugging complexity. Strategies like voting aggregate diverse outputs but add latency, while sequential routing focuses on specialized models. Effective orchestration shrinks the action space with precise definitions, limiting confusion from too many options.

Strategic Decisions on Framework Adoption and System Complexity

Direct API calls grant transparent control over the agent core, while SDKs like Strands Agents SDK wrap prompts and responses, obscuring critical debugging details. Building directly with LLM APIs often suffices, preventing unnecessary complexity. Teams should optimize single calls with retrieval and in-context examples before considering heavy frameworks.

Takeaway

Agentic systems thrive on simplicity, with workflows and agents serving distinct purposes based on task volatility. Augmented LLMs, enhanced by retrieval, tools, and memory, form the core of agentic architecture. Direct API calls and well-documented interfaces offer clearer control than opaque frameworks, ensuring maintainability and reliability in production environments.