Physical AI & Embodied Robotics
- Exponential Industry
- Technology Area
- 32 Milestones
Artificial intelligence systems and autonomous machines that operate directly in and interact with the physical world, bridging bits to atoms. Spanning foundation models (Vision-Language-Action and World Foundation Models), physics-based simulation with synthetic data (Omniverse, Isaac Sim) to safely bridge Sim2Real, and embedded runtime computers (Jetson Thor, DRIVE AGX) executing closed-loop perception-action loops across humanoids, AMRs, adaptive manipulators, autonomous mobility, and smart spaces.
Exponential Industry · Hidden Factory
Physical AI & Embodied Robotics
Universal Robots Gen 7 Platform Launch
AGIBOT GE-Act 2.0 World Action Model Release
AGIBOT released GE-Act 2.0, an embodied World Action Model pretrained across 30,000 hours of manipulation data validating scaling laws for general-purpose robot manipulation /AGIBOT / X/ /AGIBOT / YouTube/.
microagi microSLAM Monocular Spatial AI System Announcement
microagi announced microSLAM, developed with ETH Zurich CVG Lab to reconstruct dynamic 3D environments from single-RGB worker video feeds for behavioral robot learning /Gianni Hodel / microagi / X/ /microagi / YouTube/.
Skild AI Crosses $100M ARR in 10 Months from First Commercial Deployment
Skild AI achieves $100M ARR milestone within 10 months of initial deployment across 60+ customer sites, scaling factory manipulation for Foxconn/NVIDIA Blackwell racks, Sumitomo wire harnesses, and Mitsui commercial kitchens /Skild AI/.
Unitree UnifoLM-WLA-1.0 Embodied Model Open-Source Release
Unitree open-sourced the UnifoLM-WLA-1.0 foundation model for whole-body and tabletop manipulation across humanoid robot platforms /Unitree Robotics / X/ /cyberrobooo / X/.
Caterpillar and FieldAI Industrial Physical AI Partnership
Caterpillar and FieldAI partner to deploy autonomous navigation and multimodal embodied foundation models into construction, quarrying, and mining vehicles /Caterpillar Inc./.
Skild AI launches S1 Generalist Robotic Foundation Model
Skild AI released S1, a generalist robotic foundation model capable of in-context learning from a single video demonstration across 10-minute multi-step manipulation tasks with 66% zero-shot success without fine-tuning /Skild AI/ /YouTube/.
Ludo Robotics Unveils Ludi-0.1 Social Humanoid Foundation Model
Ludo Robotics, founded by Krafton, unveiled Ludi-0.1, an agentic foundation model for social humanoid robots integrating multimodal perception, speech, gestures, and manipulation tested on Unitree G1 /Ludo Robotics/.
Wang Xingxing says humanoids reach mass market in 2–10 years
At the World Robot Conference in Beijing (Thursday, Aug 20, 2026), Unitree founder Wang Xingxing said humanoids will reach mass-market popularity once they can handle 80% of voice-instructed tasks in unfamiliar surroundings, which he said will take two to ten years. The humanoid brain / embodied AI is now Unitree’s main focus /Bloomberg/.
Generalist announces GEN-1.5 one-shot embodied model
Generalist announced GEN-1.5 on Aug 19, 2026: an embodied foundation model that learns new short-horizon dexterous tasks from 3–12 seconds of a single demonstration without gradient updates (59% avg on 10 tasks) or 1–10 gradient steps on minutes of data (83%). First model the company knows of with general one-shot/few-shot closed-loop physical skill learning at this breadth /Generalist/ /X / @GeneralistAI/.
Fidler, Gojcic, and Ling launch Veeda AI for simulated-reality world models
Sanja Fidler announced Veeda AI on LinkedIn (activity id timestamp 2026-08-19T15:12:17Z), co-founded with Zan Gojcic and Huan Ling. Mission: simulated reality / world models so robots can do interactive trial-and-error at scale instead of in the real world. Investors named: Khosla Ventures and Radical Ventures /LinkedIn / Sanja Fidler/ /LinkedIn / Zan Gojcic/ /Veeda AI/.
FORT Robotics announces SPAC combination at $500M equity value
FORT Robotics and Newbury Street II (NTWO) announced Aug 18, 2026 a business combination to list as FORT Robotics Holdings (Nasdaq FROB). Pro-forma EV $556.6M; pre-money equity $500M. Expected close Q4 2026, subject to shareholder and regulatory approvals /FORT Robotics/.
Publication of Zetta ζ Closed-Loop Embodied Harness Paper
Researchers published 'Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence' on arXiv, introducing decoupled runtime critic evolution on RoboCasa and LIBERO-Pro without base policy fine-tuning /arXiv/.
Unitree $900M Shanghai IPO 8,000x oversubscribed
Reuters Aug 10, 2026: Unitree said its ~$900M Shanghai IPO was more than 8,000 times oversubscribed by retail investors. Priced at 150.80 yuan, valuing the company at more than 60 billion yuan; first onshore-listed humanoid robot maker in China /Reuters/.
robots.online 100-Robot Public Interaction Experiment
100 real AI-powered robots (hangars in Israel and California) controllable online in real time to study instinctive human-robot interaction and feed data into foundation models and interfaces.
Agility Robotics SPAC Merger with Churchill Capital Corp XI
Definitive business combination (June 24, 2026): $2.5B pre-money equity value; >$620M expected gross proceeds including ~$200M PIPE at $10/share led by Foxconn. Expected ticker AGLT; close expected 2026 subject to shareholder/regulatory approvals.
UAW convention: Fain 'fight for humanity' on AI/robots
UAW President Shawn Fain told the quadrennial convention that AI and humanoid robots are a profound threat, said 'We are in a fight for humanity,' and compared AI to NAFTA /The Detroit News/.
Robotiq IQ Platform Launch
Robotiq launched the IQ platform to automate robotic workcell engineering and validation from CAD and site scans /Business Wire/.
Advantech deploys agentic factory brain on Jetson Thor
Advantech building and deploying an agentic factory brain in its own plants with NemoClaw, Nemotron 3 and Jetson Thor (fleet management, defect detection, autonomous decisions) /NVIDIA Blog/.
Hexagon Robotics integrates Jetson Thor on humanoids
Hexagon Robotics integrating Jetson Thor and Yocto OS customization for humanoids in manufacturing, logistics and construction /NVIDIA Blog/.
Solomon coordinates humanoid agents with NemoClaw
Solomon uses NemoClaw to coordinate reasoning, perception, sensor fusion, locomotion and manipulation on a humanoid, plus Solomon active perception /NVIDIA Blog/.
Zipline uses Jetson Orin NX on delivery drones
Zipline uses Jetson Orin NX and a Yocto custom OS on autonomous drones for medical, food and retail deliveries /NVIDIA Blog/.
JetPack 7.2 lands NemoClaw agentic AI on Jetson
NVIDIA Blog 2026-06-01 (COMPUTEX): JetPack 7.2 + NemoClaw on Jetson; AGX Orin 32GB to 241 TOPS (+20%); Yocto, CUDA 13 on Orin, MIG on Thor. Customer uses include Advantech factory brain, Solomon humanoid, SandStar retail (30+ countries), Zipline drones /NVIDIA Blog/.
Inbolt Vision-Enabled Robot Programming Launch
Inbolt announced vision-enabled CAD-to-factory-floor robot programming at Automate 2026 /Business Wire/.
Physical Intelligence announces π0.7 steerable generalist
Physical Intelligence announced π0.7 on Apr 16, 2026: a steerable VLA with compositional generalization, language coaching on unseen appliances, and cross-embodiment laundry folding on a UR5e system with no laundry data for that robot. Matches specialist Recap throughput on several tasks /Physical Intelligence/.
Generalist announces GEN-1 mastery of simple physical tasks
Generalist announced GEN-1 on Apr 2, 2026: 99% average success vs 64% for prior models on the same simple tasks, ~3x faster than prior SOTA, ~1 hour of robot data per result. Pretraining uses no robot data (wearable human grippers); half a million hours of physical interaction /Generalist/.
Boston Dynamics unveils Atlas product version at CES 2026
product Atlas unveiled at Hyundai media day; production to begin immediately at Boston HQ /Boston Dynamics/.
Atlas 2026 deployments committed to Hyundai RMAC and Google DeepMind
all 2026 Atlas deployments fully committed; fleets scheduled to ship to Hyundai Robotics Metaplant Application Center and Google DeepMind /Boston Dynamics/.
Physical Intelligence announces π*0.6 Recap VLA
Physical Intelligence blog Nov 17, 2025 (arXiv Nov 18): π*0.6 trained with Recap RL from demonstrations, corrections, and autonomous experience. More than doubles throughput on some hard tasks and can cut failure rates by 2x; espresso, box assembly, novel laundry /Physical Intelligence/ /arXiv/.
Generalist announces GEN-0 robotics scaling laws
Generalist announced GEN-0 on Nov 4, 2025: embodied foundation models trained on 270,000 hours of real-world manipulation with measured scaling laws and a ~7B parameter intelligence threshold (smaller models ossify) /Generalist/.
Figure 02 factory tests at BMW Spartanburg
Figure tested Figure 02 for data collection and use-case training at BMW Spartanburg /NVIDIA Blog/.
Stellantis Detroit Inbolt 3D Vision Deployment
Stellantis deployed Inbolt GuideNOW 3D vision guidance on Fanuc robot arms at Detroit Assembly Complex, slashing downtime by up to 95% on Dodge Durango and Jeep Cherokee panel picking racks /NYT/.
Frequently Asked Questions
What is Physical AI & Embodied Robotics and what industrial engineering problems does it address?
Artificial intelligence systems and autonomous machines that operate directly in and interact with the physical world, bridging bits to atoms. Spanning foundation models (Vision-Language-Action and World Foundation Models), physics-based simulation with synthetic data (Omniverse, Isaac Sim) to safely bridge Sim2Real, and embedded runtime computers (Jetson Thor, DRIVE AGX) executing closed-loop perception-action loops across humanoids, AMRs, adaptive manipulators, autonomous mobility, and smart spaces.
What core technologies and manufacturing architectures comprise Physical AI & Embodied Robotics?
Key technical architectures and manufacturing innovations include:
- NVIDIA NemoClaw: NVIDIA agentic AI framework. Landed on the Jetson stack at COMPUTEX 2026 — single-command deploy from servers/workstations into robotics, inspection and industrial automation.
- Universal Robots Gen 7 Platform: Next-generation physical AI collaborative robotics platform with integrated multi-modal perception and real-time path planning.
- AGIBOT GE-Act 2.0 Native World Action Model: Pretrained World Action Model trained from random initialization across 30,000 hours of embodied physical data. Unifies future visual state generation with continuous robotic motor action prediction, achieving zero-shot policy generalization across 100 atomic manipulation skills and multi-robot embodiments.
- FieldAI Physical AI Foundation Models: Multimodal embodied intelligence foundation models designed for navigation, terrain perception, and autonomous operation of heavy machinery.
- NVIDIA JetPack 7.2: Jetson software stack (OS, compute, deterministic performance). 7.2 adds agentic AI skills, Yocto-based OS, CUDA 13 on Jetson Orin, a 241 TOPS AGX Orin 32GB boost, and MIG on Jetson Thor.
- GEN-0: Generalist embodied foundation model (Nov 4, 2025). Multimodal training on high-fidelity physical interaction; Harmonic Reasoning; scaling laws and a 7B ossification threshold; pretrained on 270,000 hours of real-world manipulation data growing ~10,000 hours/week. GEN-1.5 cites GEN-0 as the start of predictable robotics scaling.
Which companies and industrial facilities lead deployment in Physical AI & Embodied Robotics?
Leading industrial manufacturers, hyperscalers, and engineering operators include:
- NVIDIA Corporation: Skild AI Crosses $100M ARR in 10 Months from First Commercial Deployment (Sep 10, 2026) — Skild AI achieves $100M ARR milestone within 10 months of initial deployment across 60+ customer sites, scaling factory manipulation for Foxconn/NVIDIA Blackwell racks, Sumitomo wire harnesses, and Mitsui commercial kitchens.
- Skild AI: Skild AI launches S1 Generalist Robotic Foundation Model (Aug 25, 2026) — Skild AI launched S1, a generalist robotic foundation model capable of in-context learning. S1 can learn complex multi-step manipulation tasks lasting up to 10 minutes from a single video demonstration, achieving a 66% success rate on unseen tasks without fine-tuning or post-training.
- Foxconn: Agility Robotics SPAC Merger with Churchill Capital Corp XI (Jun 24, 2026) — Definitive business combination (June 24, 2026): $2.5B pre-money equity value; >$620M expected gross proceeds including ~$200M PIPE at $10/share led by Foxconn. Expected ticker AGLT; close expected 2026 subject to shareholder/regulatory approvals.
- Generalist: Generalist announces GEN-1.5 one-shot embodied model (Aug 19, 2026) — Our latest robot foundation model, GEN-1.5, exhibits the beginnings of that same ability: it can learn a new task in seconds, from a single example, without gradient updates or fine-tuning.
- Robotiq: Robotiq IQ Platform Launch (Jun 2, 2026) — Robotiq introduced its IQ platform, an AI-enabled solution designed to automate robotic workcell integration and shift manufacturing from manual, expert-dependent engineering toward generated workcells, starting with robotic palletizing.
- Inbolt: Inbolt Vision-Enabled Robot Programming Launch (May 27, 2026) — Introduced at Automate 2026, Inbolt’s vision-enabled robot programming allows engineers to program robots directly from CAD models with one-shot programming, dynamic tracking for moving conveyors, and real-time sub-2ms pose updates on NVIDIA Jetson AGX Orin.
What are key commercial projects and milestone achievements in Physical AI & Embodied Robotics?
Major industrial breakthroughs and commercial milestones include:
- Robotiq IQ Platform Launch (Jun 2, 2026): Robotiq introduced its IQ platform, an AI-enabled solution designed to automate robotic workcell integration and shift manufacturing from manual, expert-dependent engineering toward generated workcells, starting with robotic palletizing.
- Ludo Robotics Unveils Ludi-0.1 Social Humanoid Foundation Model (Aug 20, 2026): Ludo Robotics introduced Ludi-0.1, a foundation model and agentic system for social humanoid robots, integrating speech, gesture, multimodal reasoning, memory, navigation, and learned manipulation on physical platforms including the Unitree G1.
- Wang Xingxing says humanoids reach mass market in 2–10 years (Aug 20, 2026): The market will take off once robots can handle 80% of tasks given to them via voice instructions in unfamiliar surroundings, Wang Xingxing said in a speech at the World Robot Conference in Beijing on Thursday. That will take two to 10 years, he said.
- Skild AI Crosses $100M ARR in 10 Months from First Commercial Deployment (Sep 10, 2026): Skild AI achieves $100M ARR milestone within 10 months of initial deployment across 60+ customer sites, scaling factory manipulation for Foxconn/NVIDIA Blackwell racks, Sumitomo wire harnesses, and Mitsui commercial kitchens.
- microagi microSLAM Monocular Spatial AI System Announcement (Sep 10, 2026): microagi announced microSLAM, a monocular SLAM system built by its Zurich computer vision team with ETH Zurich's Computer Vision and Geometry Lab ranking first on the LaMaria benchmark. Running on a single RGB stream from factory workers' glasses, microSLAM turns plant walkthroughs into interactive 3D environments to pre-train behavioral robotics foundation models.
- AGIBOT GE-Act 2.0 World Action Model Release (Sep 11, 2026): AGIBOT released GE-Act 2.0, the first native World Action Model to validate pretraining and scaling for embodied AI. Trained entirely from scratch from random initialization across 30,000 hours of embodied manipulation data, task success climbs from 17.1% to 44.1% on G1-OP with zero-shot generalization across 100 atomic tasks, with cross-embodiment transfer to G2-90D.
Related Technology Ontologies
Data Center Power & Behind-the-Meter Generation
Generation and storage sited at or next to AI/cloud campuses so the load does not wait on a transmission queue. Common stacks pair gas turbines or engines with on-site BESS; some designs add behind-the-meter solar. Hyperscalers and developers announced multi-hundred-megawatt gas+BESS campuses in 2025–2026 to serve rack-scale AI halls.
Rackscale AI Accelerators
Fully co-designed rack- and POD-scale AI systems that combine accelerators, host CPUs, scale-up fabric (NVLink-class domains), and scale-out networking in one thermal/mechanical package. Lineage runs from multi-GPU HGX/DGX boxes through Grace Hopper NVL32 domains to liquid-cooled 72-GPU racks (GB200, GB300, Vera Rubin) and competing racks such as AMD Helios. NVIDIA, AMD, and OEM rack builders (Dell, Supermicro) ship the current generation for AI-factory halls.
AI Inference Silicon
Dedicated application-specific processors and accelerators engineered for high-throughput, low-latency AI model inference in data centers and edge deployments. Architectures overcome the memory wall by utilizing massive on-chip SRAM, high-bandwidth memory (HBM3e), and specialized matrix processing engines. Commercial implementations include Etched Sohu ASICs, Groq Language Processing Units (LPUs), Tenstorrent Wormhole, and Cerebras CS-3 systems.