Figure AI just signed a massive $3.5 billion check for 100,000 NVIDIA GPUs, pushing the capital requirements for humanoid development squarely into the realm of frontier language models. On the other end of the spectrum, the open-source community is churning out $1,000 bipedal builds and simulation datasets, proving you don't necessarily need a supercomputer to push physical AI forward.
Humanoid startup Figure has committed an initial $3.5 billion in compute capacity to AI infrastructure provider Nscale to secure up to 100,000 NVIDIA Vera Rubin GPUs for training its Helix vision-language-action models. Announced on Thursday, September 3, the deployment is targeted for a facility in Barstow, Texas, in the second half of 2027, with total capacity potentially expanding past $6 billion. As part of the multi-year deal, Nscale is taking an undisclosed equity stake in Figure and serving as its preferred compute provider.
Why it matters
Training vision-language-action models on continuous multimodal video and sensor streams requires infrastructure scale on par with frontier LLMs. By securing guaranteed cloud capacity backed by equity alignment, Figure insulates its training roadmap against GPU availability bottlenecks. This massive allocation underscores how scaling laws are driving physical AI capital requirements beyond standard hardware R&D into heavy datacenter commitments.
Figure CEO Brett Adcock positions the deal as a critical vertical integration step to power proprietary data and Helix model scaling. Meanwhile, industry analysts note the circular funding structure—where the compute vendor holds equity in the customer—highlights potential risks if deployment timelines lag behind capital expenditure.
Following yesterday's coverage of NC AI and LG Electronics being selected for South Korea's $1.67 billion K-On-Device AI project, we now know the startup's specific role. NC AI will supply digital twin training environments, human-to-robot motion retargeting pipelines, and proprietary world models to eliminate cloud dependencies for autonomous home humanoids.
Why it matters
Operating bipedal hardware in unstructured domestic environments requires deterministic, ultra-low-latency control that cloud architectures cannot guarantee. By pairing localized world models with custom domestic silicon, the project targets real-time trajectory planning without network latency. For hardware founders and system architects, this highlights the growing national-level emphasis on edge silicon independence for physical AI platforms.
The NC AI and LG consortium emphasizes that on-device processing is essential for safety, privacy, and real-time balance in residential spaces. Conversely, independent software developers point out that restricting inference to low-power edge chips may limit the context window and parameter size of underlying foundation models compared to cloud-hosted alternatives.
Fresh off its record-breaking Shanghai IPO surge that pushed its valuation past $62 billion, Unitree took to IFA 2026 in Berlin for the 'Robots on the Runway' showcase on Thursday, September 3. Chinese rival AGIBOT—which already commands massive domestic market share—also headlined the event, announcing its 15,000th unit production milestone as it expands partner networks across six industrial sectors in Saudi Arabia under Vision 2030.
Why it matters
As we've tracked over recent weeks, Chinese humanoid developers are rapidly moving from domestic pilot trials into international commercial distribution. Securing European regulatory clearance and Middle Eastern enterprise partnerships demonstrates that Asian supply chains are converting manufacturing volume into global market share. This deployment velocity puts pressure on Western hardware startups to demonstrate comparable commercial traction.
Unitree and AGIBOT frame their international expansion as proof that mature manufacturing ecosystems can lower unit costs for enterprise buyers. On the other hand, Western trade representatives and policy analysts note that expanding Chinese robotic deployments face growing geopolitical scrutiny and potential data-security regulatory barriers in export markets.
Earlier this week we covered Sharpa's massive funding round—which we cited at $669 million and is now reported as $668 million—alongside its automated Dairy Queen deployment in Shanghai. Valued at $3.3 billion, the company continues to showcase its dexterous robot executing a 55-step food preparation workflow using unmodified standard human equipment without store retrofitting.
Why it matters
Deploying dexterous manipulation systems directly into unmodified commercial food-service environments tests whether multi-finger robotic hands can handle long-horizon tasks under real-world operational stress. Sharpa's zero-retrofit approach eliminates expensive store redesigns, lowering adoption friction for enterprise retail operators. The massive funding round highlights strong backing from major Chinese tech conglomerates for service-sector automation.
Sharpa and its retail partners view zero-retrofit commercial deployments as essential for proving that physical AI can deliver immediate ROI in labor-constrained service roles. Independent automation consultants emphasize that maintaining strict hygiene standards and zero-downtime reliability in food preparation remains a high operational bar during sustained daily peak hours.
Building on Qualcomm's launch of the Dragonwing edge silicon at IFA this week, NEURA Robotics showcased its physical AI platform utilizing the new IQ10 chip on Thursday, September 3. Highlighting a reported $7 billion valuation and an order book exceeding $1 billion, NEURA detailed its strategy to integrate mobile robot fleets from ek robotics, Bosch Rexroth's ACTIVE Shuttle, and ADLATUS Robotics under its Neuraverse software stack.
Why it matters
Consolidating established industrial mobile robot lines under a single shared-learning software engine allows NEURA to generate immediate revenue while scaling its long-term humanoid platform. Utilizing Qualcomm's Dragonwing silicon integrates high-level spatial AI reasoning directly with real-time motor control. This multi-fleet integration strategy offers a pragmatic model for scaling robotics production in European manufacturing.
NEURA Robotics maintains that acquiring existing AMR fleets provides the immediate operational footprint required to refine its Neuraverse AI stack. Industry competitors argue that integrating disparate hardware architectures into a single software runtime introduces severe engineering overhead that could delay humanoid shipment milestones.
Axis Robotics released Axis Sim Dataset V1 on Friday, September 4, containing over 50,000 human-teleoperated simulation trajectories on a simulated Franka Research 3 arm. The Hugging Face release coincides with the $12 million seed funding round led by Hack VC that we noted previously, allowing the startup to expand its compounding physical AI data engine alongside researchers from UC Berkeley, Johns Hopkins, and the University of Michigan.
Why it matters
The release challenges the long-held assumption that physical AI models require strictly curated, expert-only demonstration data to achieve task generalization. By proving that large-scale, crowd-sourced simulation data containing noisy trajectories can effectively train vision-language-action policies like π0.5, Axis provides an open, reproducible framework for manipulation research. This lowers data-acquisition barriers for open-source policy developers.
Axis Robotics and its academic partners contend that synthetic simulation diversity is the most scalable pathway to eliminate the physical data bottleneck. Conversely, physical AI researchers caution that sim-to-real gaps in contact-rich manipulation still require real-world tactile validation that purely simulated trajectories cannot fully replicate.
Yesterday we covered Zeroth's launch of the compact Bridge desktop humanoid; today, the company officially opened applications for its Builder Program and released the initial Bridge Geek Edition batch for developer pre-orders. Built on an AI-native architecture, the open-source software ecosystem features motion-control APIs, SDKs, and a public Skill Hub.
Why it matters
High hardware costs have historically prevented independent software researchers from testing whole-body control and physical AI policies on physical bipedal hardware. Providing a desktop-scale biped with open APIs and high-torque actuators creates an affordable physical testbed for algorithm development. This open-source software approach mirrors the broader trend of hardware standardization driven by community-contributed skill repositories.
Zeroth positions Bridge as a platform to democratize physical AI development by decoupling skill creation from expensive industrial hardware. Software maintainers applaud the open APIs, though hardware engineers note that desktop-scale dynamics do not fully capture the thermal and structural challenges of full-scale humanoids.
Maker Kayden Knapik built and programmed a 3D-printed bipedal robot named 'Stompy' in three days, detailed on Thursday, September 3. Constructed for approximately $1,000 using a Bambu Lab 3D printer and Robstride 02 motors, the 6-DOF biped runs a reinforcement learning walking policy on an NVIDIA Jetson Orin Nano processor. Featuring forward-bending knees for a human-like gait, the creator released all design files and code open-source.
Why it matters
Combining consumer 3D printing with off-the-shelf brushless actuators and edge AI processors enables rapid physical iteration of legged locomotion policies for under $1,000. Open-sourcing complete CAD files and reinforcement learning deployment scripts lowers the entry barrier for student researchers and hobbyists. This agile development model accelerates community testing of novel biped kinematics without relying on expensive lab infrastructure.
Open-source robotics developers praise the project for proving that functional RL-driven bipedal locomotion can be deployed on budget hardware in days. Mechanical engineers note that while 3D-printed structures are ideal for rapid prototyping, long-term durability and torque density still favor metal chassis for industrial deployments.
Developer Sourabh Warghane published an open-source Autonomous Mobile Robot project on Thursday, September 3, built around an NVIDIA Jetson Orin Nano 8GB and ROS 2 Humble. The robot integrates YOLOv8n-Pose TensorRT vision processing at 25-30 FPS, RPLIDAR C1 obstacle avoidance, and Faster-Whisper voice interaction powered by NVIDIA NIM. The entire software stack is fully containerized via Docker, with source code, ESP32-C3 motor controller firmware, and architecture docs uploaded to GitHub.
Why it matters
Containerizing real-time ROS 2 nodes alongside TensorRT inference pipelines simplifies software deployment and eliminates dependency conflicts for mobile robotics builders. By integrating visual pose estimation, LIDAR navigation, and local voice command processing on a single edge board, the open-source blueprint provides a modular reference architecture for autonomous service robots.
Community developers praise the containerized Docker setup for making complex ROS 2 and TensorRT environment configuration reproducible. Embedded system architects note that running concurrent vision, LIDAR, and speech models requires careful thermal management and power allocation on entry-level Jetson hardware.
Spatial intelligence startup World Labs, co-founded by Fei-Fei Li, launched Atlas on Friday, September 4, a multimodal autoregressive diffusion transformer designed for 3D generation, reconstruction, and simulation. The omni model natively processes text, images, video, and 3D data, generating up to one minute of 1440p video and reconstructing physical environments into point clouds or Gaussian splats from minimal photo inputs to power 'Real-to-Sim' robotics pipelines.
Why it matters
Generating physical environments directly from sparse 2D imagery allows robotics teams to procedurally build high-fidelity simulation environments for reinforcement learning without manual 3D asset modeling. By unifying visual generation and explicit spatial geometry into a single model, Atlas addresses the data-scarcity issue in physical AI training. This accelerates sim-to-real transfer workflows across industrial and warehouse manipulation.
World Labs asserts that integrated spatial world models offer a far more scalable pathway for synthetic data generation than traditional physics engines. However, independent robotics evaluators note that synthetic Real-to-Sim environments must still prove they can accurately model subtle physical dynamics like friction, compliance, and material deformation.
Medtronic entered a strategic partnership on Thursday, September 3, making an approximately $700 million investment in Hong Kong-headquartered Cornerstone Robotics. The deal grants Medtronic distribution rights for Cornerstone's Sentire surgical system in select international markets outside the U.S. where it holds regulatory clearance, including the EU CE Mark, Singapore HSA approval, and China NMPA clearance. Medtronic will position the vertically integrated Sentire system alongside its proprietary Hugo robotic-assisted platform.
Why it matters
Single-digit global adoption rates in robotic surgery have prompted industry leaders to adopt multi-platform commercial strategies rather than relying solely on proprietary hardware. By acquiring immediate distribution rights to an established, multi-specialty system, Medtronic expands its international market coverage without waiting for internal R&D cycles. The influx of capital validates Cornerstone's vertically integrated hardware and imaging manufacturing model.
Medtronic leadership views the partnership as a capital-efficient method to scale procedural volume globally while offering hospital systems versatile robotic options. Sector analysts observe that success will depend on navigating fragmented international reimbursement structures and competing against entrenched incumbents like Intuitive Surgical.
Xiaomi unveiled three custom silicon additions to its Xuanjie chip family on Friday, September 4: the 3nm Xuanjie O3 mobile SoC, the Xuanjie O100 high-bandwidth AI accelerator, and the 3nm Xuanjie D100 intelligent driving processor. The O100 uses 6nm 3D wafer-level stacking to achieve 1.22 TB/s of memory bandwidth, while the D100 supports up to 160 GB of unified memory and models up to 200 billion parameters for autonomous vehicle and edge AI deployment.
Why it matters
Consumer electronics and EV giants are increasingly engineering custom silicon to bypass third-party chip shortages and eliminate memory bandwidth bottlenecks in edge inference. Utilizing 3D wafer-level stacking provides the bandwidth necessary to run large multimodal perception models locally. This custom silicon foundation supports Xiaomi's broader ecosystem spanning autonomous driving, smart home hardware, and robotics.
Xiaomi executives frame proprietary silicon as a necessary foundation to deliver tight hardware-software integration across their product lines. Semiconductor market analysts note that high tape-out and packaging costs for 3nm and 3D-stacked chips require massive shipment volumes across mobile and automotive lines to justify the investment.
Ezurio announced the Carbon AM67 System-on-Module on Thursday, September 3, built on the Texas Instruments AM67x processor family and an integrated Wi-Fi 6 module. Measuring 45 x 30 mm in an OSM-MF form factor, the module features four Arm Cortex-A53 cores, three Cortex-R5F real-time MCU cores, and two deep learning accelerators providing 4 TOPS of edge AI compute. It incorporates Ezurio's EZ BSP framework with a hardware root of trust for embedded security.
Why it matters
Industrial robotics OEMs require compact computing modules that combine neural network acceleration with dedicated real-time microcontrollers for motor control. Pairing Arm Cortex-A cores with Cortex-R real-time units on a solderable OSM form factor reduces carrier board complexity. This enables rapid commercial development of secure, ruggedized edge controllers for mobile robots and medical devices.
Ezurio emphasizes that combining TI processing silicon with hardware-enforced root-of-trust security simplifies cyber-resilience compliance for industrial OEMs. System integrators note that while 4 TOPS is suitable for localized vision and sensor processing, heavy vision-language workloads will still require external accelerator co-processors.
MIPS, operating under GlobalFoundries, launched three physical AI developer platforms on Thursday, September 3: Acies, Actus, and Aegis. Targetting edge inference, real-time event processing, and safety-critical robotics control, the platforms combine open RISC-V architectures with software-first workload modeling. GlobalFoundries provides manufacturing scale and advanced packaging to transition MIPS from IP licensing to full system-level silicon solutions.
Why it matters
The transition of RISC-V from basic microcontrollers into domain-specific, safety-certified physical AI accelerators provides robotics builders with an open alternative to proprietary instruction sets. Direct manufacturing backing from GlobalFoundries ensures enterprise supply chain reliability for industrial and automotive OEMs. Having dedicated hardware tiers for real-time safety control versus inference simplifies system integration for complex autonomous platforms.
MIPS and GlobalFoundries state that open-standard RISC-V architectures provide the customizability needed to solve domain-specific physical AI bottlenecks. Industry skeptics contend that RISC-V software ecosystem tools must continue maturing to compete directly with established ARM and NVIDIA CUDA development environments.
Yesterday we covered Pollen Robotics surpassing 10,000 pre-orders for its $399 Microduck bipedal robot. While earlier reports noted $5 million in total revenue, the company highlighted generating over $1 million within the first six hours of launch. To meet the massive developer demand, volume production is being handled in partnership with Seeed Studio in Shenzhen.
Why it matters
High pre-order volume for a sub-$400 biped demonstrates strong developer demand for affordable, physical reinforcement learning hardware. Partnering with Seeed Studio in Shenzhen allows the startup to leverage mature consumer electronics supply chains without building proprietary manufacturing facilities. This establishes a clear commercial blueprint for democratizing desktop robotics platforms.
Pollen Robotics highlights the rapid order volume as validation that open-source software paired with low-cost bipedal hardware hits a major market sweet spot. Manufacturing analysts point out that scaling from 10,000 pre-orders to reliable volume delivery will test Seeed Studio's quality control across multi-actuator assemblies.
Segway Navimow introduced its H510 Pro robotic lawn mower at IFA 2026 on Friday, September 4, equipped with a motorized mechanical arm that extends up to 2.9 inches for zero-gap border trimming. The all-wheel-drive mower features crab-walking 'Xero-Turn' wheel rotation to prevent turf tear during tight maneuvers, EdgeSense surface recognition software, and thermal heat-sensing safety shut-offs to detect animals or humans near the blades.
Why it matters
Perimeter trimming has remained a primary manual maintenance task that boundary-wire and RTK robotic mowers failed to eliminate. Integrating an extending mechanical arm onto an outdoor mobile chassis directly solves edge-coverage limitations. Adding active thermal sensing and crab-walking drive architectures raises the hardware capability standard for consumer lawn robotics.
Segway Navimow asserts that extending mechanical arms and active thermal detection eliminate manual touch-up trimming while improving yard safety. Consumer hardware reviewers point out that added mechanical complexity in outdoor environments increases potential wear points from dirt, grass buildup, and moisture exposure.
Researchers introduced Hydraulically Amplified Low-Voltage Electrostatic (HALVE) actuators in Science Advances on Friday, September 4, matching mammalian skeletal muscle in average power density (50.5 W/kg) and peak strain rate (971%/s). Operating at 1,100 volts—a 4.9-fold reduction compared to prior electrohydraulic designs—the composite actuators combine a PET shell with a high-permittivity PVDF terpolymer layer. The waterproof, self-clearing actuators were demonstrated on a robotic gripper and soft swimmer.
Why it matters
Electrohydraulic soft actuators have historically required bulky, dangerous multi-kilovolt power supplies that made untethered operation impractical. Dropping driving thresholds to 1,100 volts allows system designers to utilize standard, low-cost commercial MOSFET semiconductor switches instead of specialized optocouplers. This voltage reduction opens pathways for lightweight, untethered soft robotics and multi-channel wearable devices.
The research team highlights that HALVE actuators eliminate heavy power-tether constraints while maintaining high force output and self-healing durability. Soft robotics engineers note that while 1,100V is a major improvement, further voltage reductions will be necessary to meet strict consumer wearable electrical safety standards.
Following our coverage yesterday of Shenzhen Hong Ye Jie's new platinum silicone skin announced on September 2—now specified as the HYJ-RS-033 series rather than the previously cited HYJ-RS-035—the company is emphasizing its zero-bleed, dry-touch properties. Utilizing high-purity platinum addition curing, the material prevents surface stickiness and dust accumulation over tens of thousands of bending cycles for dynamic humanoid joints.
Why it matters
Dynamic silicone skins on commercial humanoids frequently degrade due to oil precipitation, permanent creasing, and surface dust collection during continuous operation. Formulating fatigue-resistant, zero-bleed platinum silicone solves a major physical durability bottleneck for tactile interfaces and facial expressions. High elasticity across high-movement joints extends maintenance intervals for deployed service humanoids.
The material developer emphasizes that zero-bleed platinum curing eliminates surface degradation and sticky residues in long-term field deployments. Roboticists note that while dynamic skin materials are improving, integrating soft skins with underlying tactile sensor arrays without dampening touch sensitivity remains an engineering challenge.
Researchers in Argentina detailed spray-dried whey protein–chitosan microcapsules on Friday, September 4, engineered to shield probiotic bacteria through gastric transit. Measuring 2 to 15 micrometers, the microcapsules increased bacterial adhesion to intestinal mucus by 77-fold and boosted survival 100-fold during simulated digestion. Once anchored, the encapsulated Lacticaseibacillus casei BL23 continued secreting 70-100 nm anti-inflammatory extracellular vesicles directly to intestinal tissue.
Why it matters
Oral delivery of microscale biological therapeutics typically fails due to gastric acid destruction and rapid intestinal clearance. Utilizing protective polymer microcapsules as anchored in-situ bioreactors solves both transit survival and local mucosal retention. This provides a scalable micro-encapsulation strategy for targeted drug delivery and bio-hybrid medical robotics.
The research team highlights that in-situ vesicle production eliminates the need to isolate and stabilize fragile extracellular vesicles prior to administration. Biomedical engineers note that moving from simulated digestive models to live animal trials will be required to confirm therapeutic efficacy and uniform mucosal coverage.
Autonomous trucking developer PlusAI announced a definitive SPAC business combination with Texas Ventures Acquisition III Corp on Thursday, September 3, valuing PlusAI at $800 million pre-money. The deal is expected to deliver up to $300 million in gross proceeds, including $60 million in committed financing from Yorkville Advisors. The capital will fund the 2027 commercial rollout of its Level 4 SuperDrive system across factory-built truck lines with partners TRATON, Hyundai, and IVECO.
Why it matters
Securing public market funding via a SPAC injection provides PlusAI with necessary capital to complete long-horizon OEM integration and regulatory validation through 2027. Partnering directly with major truck OEMs for factory installation avoids expensive aftermarket retrofits. Concurrently monetizing its HyperFoundry AI simulation platform provides near-term software revenue to offset autonomous R&D burn rates.
PlusAI leadership emphasizes that factory-built OEM integration paired with dual software licensing provides a capital-efficient path to scale L4 freight. Financial analysts note that SPAC transactions carry high redemption risks, requiring committed financing pipelines to ensure projected capital targets are actually realized.
Compute Infrastructure Scale Escalates for Physical AI Models Humanoid developers are locking down dedicated multi-billion-dollar cloud infrastructure and tens of thousands of next-generation GPUs. Processing continuous multimodal sensor streams and executing reinforcement learning in simulation now rivals or exceeds the compute intensity of large-language model pre-training.
Open-Source Hardware Kits Lower Bipedal Prototyping Thresholds Community-driven open hardware releases and low-cost desktop biped platforms are making simulation-to-real reinforcement learning accessible to independent researchers. Desktop humanoids and 3D-printed biped designs are bridging the gap between theoretical software policies and physical execution.
Strategic Alliances Accelerate Global Surgical Robotics Footprints Major medical device conglomerates are executing multi-hundred-million-dollar strategic distribution agreements and investments with international startups. Partnering with localized developers allows established players to bypass lengthy regulatory cycles and deploy multi-platform surgical options globally.
Domain-Optimized Silicon Accelerates Edge Vision and Control Custom semiconductor designs and heterogeneous SoCs are prioritizing high-bandwidth local memory and deterministic control loops over raw cloud connectivity. Integrated edge platforms are enabling low-latency spatial reasoning and motor control directly within localized robotic enclosures.
Material Innovation Yields Compliant, Fatigue-Resistant Actuators Research and commercial releases in anti-swelling hydrogels, stretchable polymers, and zero-bleed silicones are solving long-standing physical wear bottlenecks. These compliant materials enable dynamic robot skins and soft actuators to endure thousands of motion cycles without structural failure.
What to Expect
2026-09-14—Culver City Council holds public hearing on proposed zoning restrictions for autonomous vehicle fleet facilities.
2026-09-14—ECOVACS limited-time launch promotional pricing concludes for WINBOT W2S PRO OMNI window cleaner.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
519
📖
Read in full
Every article opened, read, and evaluated
144
⭐
Published today
Ranked by importance and verified across sources
20
— The Robot Beat
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste