Topic

embodied ai

9 signals

Infrastructure2 hours ago

Building Fixed Hardware Implementations of Neural Networks

Researchers from Yale, Cornell, Boston University, and NTT Research published "Physical Foundation Models: Fixed hardware implementations of large-scale neural networks," arguing that the ~1-year release cadence of trillion-parameter foundation models justifies building fixed, special-purpose hardware where the network is realized directly in physical structure (such as 3D nanostructured glass) rather than executed on general-purpose digital inference chips.

Why it matters

The economic argument here matters more than the optics demo. The authors are saying that once a single foundation model gets reused across thousands of downstream tasks, the amortization math flips: it becomes rational to tape out hardware whose weights are physically baked in, on the same yearly cadence as model releases. That is a direct challenge to the Nvidia general-purpose GPU thesis for inference, and it is the same logic Groq, Etched (Sohu), and MatX are already pursuing in digital silicon. PFMs push it one step further into analog physics.

For operators, the near-term read is not glass holograms running GPT-class models. It is edge inference: trillion-parameter capability in a power envelope a robot or drone can carry, which today caps out at far smaller models. If even a fraction of the claimed energy-efficiency gain materializes in nanoelectronic form, the on-device vs cloud calculus for humanoids and AVs changes. Watch NTT Research specifically; they have been quietly funding coherent optical compute for years and are the most likely party to move this from arXiv to a fab partnership.

Economics3 hours ago

Software becoming the biggest bottleneck to physical AI innovation, finds QNX research

Source: The Robot Report

QNX, the BlackBerry-owned embedded OS division, released an OnePoll survey of 1,000 robotics software developers finding 27% cite software architecture and integration as their biggest performance bottleneck versus 16% for hardware, while 91% still run real-time workloads at least partly on general-purpose operating systems. Two-thirds of respondents reported certification-related project delays, rising to 70% in the UK and Germany versus 56% in China.

Why it matters

The number to dwell on is the certification delay gap: 70% of UK and German teams report project slippage from certification versus 56% in China. That is a 14-point speed advantage that compounds across product cycles, and it is showing up in the survey data before it shows up in deployment counts. For anyone modeling time-to-revenue on a humanoid or mobile manipulator, the regulatory cost of operating in Europe is now quantifiable and material.

The other tell is that 91% of developers admit they are running deterministic, safety-critical workloads on general-purpose operating systems (read: Linux), and 86% of that group say they are open to switching. That is a vendor's dream survey, and QNX commissioned it, so discount accordingly. But the underlying tension is real: the robotics stack was built on ROS plus Ubuntu because it was fast to prototype, and nobody wants to be the company that ships a hospital robot certified on a kernel they cannot freeze. Expect a wave of safety-certified RTOS procurement over the next two years, with QNX, Wind River VxWorks, and Green Hills INTEGRITY as the obvious beneficiaries.

Economics6 hours ago

This AI startup will clean your home for free to train future robots

Source: The Verge

Shift, an AI training data startup, is offering free home cleanings in New York in exchange for first-person camera footage captured by a camera-equipped hat worn by the cleaner, with the data used to train household robots. The company says expansions to San Francisco, London, Zurich, and Munich are coming "very soon," and that it already pays tens of thousands of people across 15 countries to record activities through its app.

Why it matters

The interesting number Shift is not disclosing is the implicit price per hour of egocentric household data. A NYC deep clean runs roughly $150 to $250 for two to three hours; if Shift is willing to eat that cost in exchange for ego-view footage in dirty, cluttered, real homes, that puts a floor on what household manipulation data is worth right now to whoever is buying. For reference, Scale and Surge pay $20 to $40 per hour for staged kitchen demonstrations. Shift is paying roughly 3 to 5x that for unstaged environments, which tells you where the bottleneck actually is for humanoid and home-robot training: not annotation, but distribution into private homes.

The structural play is the recruiting funnel. Shift already pays "tens of thousands" of contributors across 15 countries through its app, and the cleaning offer is a customer acquisition channel for households, not cleaners. Operators building home robots (1X, Figure, Weave, Physical Intelligence) need exactly this kind of long-tail residential data and have no clean way to source it themselves without consent headaches. Watch whether Shift signs an exclusive with a humanoid OEM or stays horizontal like Scale did. The plumbing, cooking, and construction expansion telegraphs the latter.

InfrastructureMay 28

NVIDIA Research Advances Robotics From Simulation to the Real World

Source: NVIDIA Blog

NVIDIA Research presented 28 papers at ICRA, with eight focused on simulation-to-real transfer across multi-arm scheduling (ScheduleStream, 3x speedup), cross-embodiment navigation (COMPASS, ~80% success on real robots), grasping in clutter (Grasp-MPC, 75% vs 41% baseline), precision assembly (SPARR, 38% success improvement; Refinery, 91% sim success), and vision-language-action grounding (PEEK, up to 41x accuracy gain; SEAL, 15% gains). NVIDIA also said its Physical AI Dataset has passed 15 million downloads, and roughly 50 ICRA papers from CMU, ETH Zurich, MIT and UT Austin cite NVIDIA-accelerated simulation or compute.

Why it matters

Strip away the paper-by-paper benchmark recitation and the operator read is straightforward: NVIDIA is locking in Isaac Lab, Isaac Sim and the Physical AI Dataset as the default substrate for robot policy training, the same way CUDA locked in GPU compute. Fifty ICRA papers from CMU, ETH, MIT and UT Austin referencing NVIDIA-accelerated stacks is the leading indicator. The next generation of robotics hires arrives already fluent in Isaac, which is how a tools moat compounds.

The technical numbers worth tracking are the sim-to-real deltas, not the absolute success rates. Grasp-MPC going from 41% to 75% on novel cluttered grasps, SPARR cutting cycle time 30% on NIST assembly, PEEK delivering a 41x accuracy multiplier on sim-trained policies: these are the levers that move warehouse and assembly economics from "demo" to "billable." For operators evaluating whether to in-house policy training or buy from a foundation-model vendor like Physical Intelligence or Skild, the question is whether these techniques ship as reusable primitives in Isaac or stay locked in research repos. The GitHub release for ScheduleStream suggests the former, which would further pull policy R&D onto NVIDIA's rails.

DealsMay 26

This startup is betting India’s gig economy can train the world’s robots

Source: AI News & Artificial Intelligence | TechCrunch

Human Archive, a Silicon Valley startup founded by UC Berkeley and Stanford researchers, raised $8.2 million from Wing Venture Capital, NVP Capital, Y Combinator, and angels from OpenAI, Nvidia, Google, Mercor, and Meta to pay Indian gig workers $1/hour to wear camera caps, tactile gloves, and motion-capture rigs that collect egocentric training data for robotics labs. The company says it has more than 1,000 active headsets and 50+ devices deployed across home-services, hotel, and restaurant partners in India, with early expansion into Southeast Asia and the US.

Why it matters

The cap table is the story. Angels from OpenAI, Nvidia, Google, and Meta backing a data-collection arbitrage tells you the frontier labs view egocentric human-task data as the binding constraint on humanoid policies, and they are willing to fund $1/hour Indian gig labor to get it before the supply tightens. Human Archive's edge is not the cap. It is the claim of synchronized RGB-D plus tactile force plus full-body mocap plus wrist cameras at scale, which is the sensor stack labs actually need to train manipulation policies, not just video.

The fragility is the partnership layer. Urban Company and Pronto publicly walked, India's MeitY is now examining consent mechanisms for exactly this category, and competitors are paying workers two to four times more per hour. If DPDP enforcement tightens or a larger home-services platform launches a captive data program (Pronto is already shopping one), Human Archive's labor-cost advantage compresses fast. For operators buying physical AI training data, the question to ask vendors in 2026 is not hours collected but sensor-synchronized hours under defensible consent.

ChinaMay 25

How Xiaomi’s push into AI, chips and EVs is future-proofing its hardware empire

Source: Tech - South China Morning Post

Xiaomi committed more than 60 billion yuan ($8.8 billion) to AI investments over the next three years, and its newly released MiMo-V2.5-Pro model was ranked the top open-source system for agentic capabilities by benchmark platform Artificial Analysis. The model placed second and third in overall intelligence and coding among open-source systems, behind DeepSeek and Moonshot AI.

Why it matters

The $8.8 billion line is the one that matters for operators. Xiaomi is the only Chinese consumer hardware company shipping at scale across smartphones, smart home, and EVs (SU7 deliveries are real), and it is now funding an in-house frontier model program to sit underneath all of it. That changes the procurement calculus for anyone building agentic features on Qwen or DeepSeek: a vertically integrated competitor with its own silicon roadmap, its own vehicles, and its own distribution can subsidize the model layer indefinitely.

The quieter read is what MiMo signals about the open-source tier in China. DeepSeek, Moonshot, and now Xiaomi are clustered at the top of the agentic benchmarks, all open-weight, all Chinese. For Western robotics and AV teams building on open models, the dependency map is consolidating onto Chinese labs faster than the export-control conversation has caught up to. The interesting question for 2026 is whether Xiaomi ties MiMo directly into the SU7 and YU7 cabin stack, which would be the first time a frontier open model ships as standard equipment in a car.

BonusMay 23

The future of physical AI isn’t humanoid; it’s task-specific and cost-efficient

Source: The Robot Report

Hailo's vice president of physical AI, Yaniv Sulkes, argues in a Robot Report contributed piece that task-specific robots running edge AI will scale faster than general-purpose humanoids, citing hardware, dexterity, energy, and cost constraints. He points to deployments like Husqvarna's AI-enabled robotic lawn mowers, which use Hailo edge processors for on-device sense-think-act loops, as the template.

Why it matters

This is a Hailo positioning piece, and the cap-table read matters: Hailo sells edge inference chips and loses if the humanoid narrative consolidates compute onto Nvidia Jetson-class modules or cloud-tethered VLAs. The argument is still directionally right for operators sizing 2026 deployments. The unit volumes already shipping (robotic vacuums, lawn mowers like Husqvarna's, last-mile delivery, ag sprayers) all clear the bar Hailo describes: bounded task, embedded inference, no cloud in the control loop. Humanoid pilots are not in that category yet.

The number to watch is bill-of-materials. A task-specific robot built around a sub-$50 edge accelerator and a fixed-function actuator stack pencils out at consumer or light-commercial price points. A humanoid with 40-plus degrees of freedom and a 100W-class compute module does not, regardless of how good the foundation model gets. Operators evaluating physical AI vendors should ask where inference runs, what the per-unit silicon cost is, and whether the control loop survives a dead network. Those three questions separate shipping product from demo reels.

LaunchesMay 22

We tried Google’s AI glasses and they’re almost there

Source: AI News & Artificial Intelligence | TechCrunch

At Google I/O, Google demoed prototype Android XR glasses with an in-lens display running Gemini for translation, turn-by-turn navigation via Google Maps, photo capture, object recognition, and configurable widgets. \

DealsMay 21

Hark Raises $700M at $6B Valuation for Figure Founder's AI Hardware Venture

Source: TechCrunch

Brett Adcock's AI hardware startup Hark raised over $700 million in a Series A round valuing the company at $6 billion. Hark is building a vertically integrated family of AI hardware devices with co-developed models and interfaces; notably, Hark's models are already being trained on Adcock's Figure AI robots, creating a cross-company embodied-AI data loop.

← All signals