#1 Unitree goes public, Jetson Orin Nano 2 launches, and an 8.64s robot sprint

★ Read this first

Not one story in this issue is a new robot brain. No VLA dropped. No world model shipped weights. The seven days from August 19 to 26 produced something else entirely.

A share price. An edge module. A search index over robot recordings. A foundation model trained on metal cutting. A cleaning fleet with hundreds of machines already in buildings. Every one of them is scaffolding — the parts that decide whether a policy ever runs a shift.

That is the theme, and Beijing supplied the proof by contrast. Peak locomotion improved four times in five days. The ability to decelerate did not improve at all. The body is ahead of the stack, and this week the stack started raising money.

460%

Unitree day one

2×

inference at the edge

8.64s

robot 100 m final

01 Unitree · Public markets

Unitree Robotics surges in Shanghai share debut, the first listed pure-play humanoid maker

Until August 19 every humanoid valuation in the world was a private mark set by a handful of people in a room. Figure at $39B, Skild at $14B, Physical Intelligence at $5.6B — all negotiated, none traded. Unitree listed on Shanghai’s STAR Market and the number stopped being an opinion.

Priced at 150.80 yuan, the stock opened at 1,100 yuan — up 629% — then gave most of it back to close at 845 yuan, a 460% first-day gain and roughly 342 billion yuan of market value. The raise was about 6.1 billion yuan, near $904M, with DeepSeek in as a strategic investor. Retail demand ran past 8,000× the allocation on a day the CSI 300 fell about 3%. Read the close, not the open: the pop is the headline, the settle is the price.

Read on SCMP →  ◆ Bloomberg  ◆ Reuters figures  ◆ Forbes comparison

02 NVIDIA · Edge compute

NVIDIA announces Jetson Orin Nano 2 robotics computer to redefine entry-level edge AI

Frontier-class reasoning on a robot has meant a Thor-class module and a power budget to match. That is affordable on a humanoid and absurd on a delivery drone or a floor cleaner. The interesting move in edge silicon is not the top of the line — it is the bottom of the line catching up to last year’s top.

Jetson Orin Nano 2 is 78 TOPS with 8GB of memory and an 8-core Arm CPU in the same footprint as its predecessor. NVIDIA puts it at 2× the inference of Jetson Orin Nano Super via better Tensor Cores and memory bandwidth, and at 15W it draws 40% less power for the same work. Those are vendor figures, not third-party ones. Cognex, Doosan Bobcat and Matic are named as first adopters; Alphabet’s Wing runs the previous module in its delivery fleet and says it will evaluate this one.

NVIDIA newsroom →  ◆ Robotics 24/7  ◆ Jetson AI Lab  ◆ Isaac GR00T

03 Foxglove · Robot data

Introducing the agentic data platform for physical AI

Finding the moment a robot failed is still a person scrubbing a timeline, writing a throwaway script, and tabbing through plot fields. The recordings are there. The index over them is not. Foxglove built the index, and then put an agent in front of it.

Semantic Search runs vector search over frame embeddings: on ingest it decodes image and video topics, samples one frame per topic per second, embeds them, and stores the vectors in Lance tables in object storage. Device, time range and topic filter the candidates first; cosine similarity ranks what is left. Object and scene search works inside one frame — action search reasons across sequences using NVIDIA Cosmos open world models, which is what catches motion no single image contains. Images and video only for now; other robot data types are stated as the goal, not the shipped state.

Foxglove blog →  ◆ Semantic Search  ◆ Agent Sidebar  ◆ Actuate 26 coverage

04 Limitless Labs · Foundation models

Limitless Labs raises $20M Series A to expand its physical AI foundation model for precision manufacturing

Almost every physical AI foundation model is trained on video of bodies moving through space. This one is not. Limitless Labs, formerly LimitlessCNC, trained its model on the physics of metal cutting, CAD geometry and the operating constraints of specific machines — a narrow world where the ground truth is a finished part.

The $20M round is co-led by Dell Technologies Capital and Square Peg, with Grove Ventures, Meron Capital and Kinetica. The model drives a CAM Agent that lives inside Mastercam, NX and Creo: given a CAD file it identifies features, picks tools, sequences operations and generates toolpaths, with the engineer keeping control of the workflow. The Israel-based company reports production deployments with Sandvik and Iscar and cites up to 50% less CNC programming time — its own figure, and an upper bound rather than a typical one. Platform is ITAR-compliant and runs on AWS GovCloud. Closed-loop CNC automation is the roadmap, not the product.

Announcement →  ◆ CAM Agent  ◆ Robotics 24/7

05 NEURA · Service robots

NEURA acquires ADLATUS Robotics, bringing physical AI into professional cleaning

The hard part of a service robot business is not the robot. It is the hundreds of buildings that already let one in, the navigation stack tuned to their floors, and the contracts that renew. NEURA bought all three rather than build them, taking 100% of ADLATUS of Ulm — its fifth acquisition in sixteen months, after Bosch Rexroth’s ACTIVE Shuttle.

ADLATUS brings autonomous cleaning and sweeping robots across industry, logistics, healthcare and public spaces, with hundreds of systems in the field and its own navigation software. The plan is to add sensing and AI and connect the fleet to Neuraverse, so a machine reads surfaces and soil types and picks a strategy instead of running a fixed route. The two have partnered since September 2025. Terms were not disclosed, and no timeline was given for when any of the AI capability actually lands on the installed base.

NEURA announcement →  ◆ Tech.eu  ◆ Deal context  ◆ Robotics 24/7

One more thing…

Beijing · Locomotion

Tiangong Ultra resets the 100 m mark at 8.64 s as the humanoid robot games close in Beijing

You have watched the clip by now. What the clip does not show is the shape of the week. Tiangong Ultra ran 9.39 s in an opening heat on Saturday, 8.86 s in the semi-final, and 8.64 s in Wednesday’s final — four times the mark fell in five days. A year ago the same event was won in 21.50 seconds.

The rest of the scoreboard reads the same way. 400 m in 38.15 s against a human record of 43.03 s. A standing high jump of 3.4 m, up from 2.88 m earlier in the week and from roughly 0.95 m for the best robot last year. Honor’s Lightning went 8.94 s in the same semi-final. 666 teams and 2,056 robots across 51 events, from a machine built by the Beijing Humanoid Robot Innovation Center, also known as X-Humanoid.

Then the part that matters to anyone shipping a policy. After the 8.86 s run the robot could not decelerate, hit the padded stopping mat, collapsed, and a small fire appeared in its torso. In the other semi-final a competitor came apart mid-race. Acceleration is solved well enough to embarrass a world record. Controlled stopping, in a space shared with people, is not. One team’s runner also finished with its arms held near its face — reinforcement learning in simulation found that hip rotation plus that posture beat imitating a human arm swing, which is a genuinely useful result buried under the spectacle.

These are competition times under competition rules, not World Athletics records, and the comparison to Bolt is a benchmark rather than an equivalence. Take them as what they are: a body that improved four times in a week, waiting on a stack that did not.

Global Times final →  ◆ Al Jazeera  ◆ The gait result  ◆ Daily briefs

Also worth the click

Intel’s Robotics Readiness Survey — 60% of leaders say they are betting big on robotics, 40% say they are ready to scale it. The gap is strategy, skills, safety and infrastructure.

Physical AI on Marktechpost — our running coverage of robot foundation models, world models and VLAs, including the Dyna-2 world-action model and Gemini Robotics 2 breakdowns.

Sponsor AI Dev Signals

Reach the people putting models on hardware, not the people posting about it.

This list is engineers, researchers and founders who read a model card and a datasheet before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

The Physical AI edition ships with four placements per issue. All four are open.