AI robotics is the intersection of modern AI (transformers, vision-language-action (VLA) models, foundation models) and physical robotics. It is distinguished from traditional robotics control in that AI-era robotics treats perception, planning, and actuation as a unified foundation-model problem rather than as separately engineered modules. The field spans humanoid robots, foundation models for embodiment, autonomous vehicles (AVs), and drones, and in 2026 has converged on questions of how to generate the simulated dynamics and real-world action data needed to train embodied systems at scale.
World models as proposed enabling technology
A central architectural debate in robotics during 2026 concerns world models — interactive predictive models that simulate spatial-temporal environments in response to actions ($P(s_{t+1} \mid s_t, a_t)$), distinct from LLMs that predict $P(x_{t+1} \mid x_t)$ over tokens and from video models that predict the same over pixels.
McCormick & De Witte (Not Boring, March 2026) argue that traditional simulation engines scale at $O(N)$ to $O(N^2)$ in the number of agents, objects, and interactions, which makes robotics-grade real-world simulation intractable. World models collapse this to a fixed-cost forward pass through a neural network whose weights have absorbed real-world dynamics from training video. McCormick and De Witte frame this as the proposed solution to a long-standing robotics bottleneck.
Three approaches to world models were active in 2026:
- Generative pixel-space — World Labs, Google DeepMind, and NVIDIA Cosmos.
- Latent / JEPA — Yann LeCun's AMI Labs ($1.03B raised).
- Action-conditioned — General Intuition ($133.7M Seed), trained on observation-action triples from gaming clips.
Several 2025–2026 deployments and research efforts have been cited as proof points for the world-models approach in robotics. Ai2's MolmoBot (2025) is an open world-model suite for robotics that demonstrated sim-to-real zero-shot transfer for manipulation. Tutor Intelligence's "Data Factory 1," unveiled by Josh Gruenstein in May 2026, is a 100-robot teleoperation facility for scaling industrial-semi-humanoid training data.
The Niantic Spatial / Coco Robotics partnership (March 2026, MIT Technology Review) is a deployment of a location-precision spatial model for last-mile delivery. Niantic's visual positioning model is trained on 30 billion crowdsourced Pokémon Go and Ingress images and pinpoints location to within centimeters from snapshots, addressing GPS failure in dense urban canyons where GPS drift of roughly 50 meters is typical. The model is deployed on Coco's 1,000+ delivery robots in Los Angeles, Chicago, Jersey City, Miami, and Helsinki. Niantic Spatial CEO John Hanke said: "It turns out that getting Pikachu to realistically run around and getting Coco's robot to safely and accurately move through the world is actually the same problem."
The data layer for physical AI
A competitive layer has formed around where humanoid-robot training data comes from. The contest is between robot teleoperation, in which operators puppet robots to generate demonstrations (for example, Tutor Intelligence's 100-robot "Data Factory 1"), and human-sourced motion capture, which records how people actually move and folds that into training.
Mecka AI (New York) disclosed in June 2026 that it had raised $60 million — a $25 million Series A closed in November plus a $35 million follow-on — led by Framework Ventures, with participation from Menlo Ventures, SV Angel, and Kindred Ventures. The company captures human-movement data from body sensors and iPhones to train humanoid robots. CEO Josh Gao projected a $100 million annual run rate from already-signed contracts and argued that human-sourced motion data, rather than robot teleoperation, becomes the defensible data layer of physical AI. (Source: fortune.com)
This poses a question parallel to the world-models debate: whether the binding constraint on humanoids such as Figure's and Physical Intelligence's embodiment-flexible models is the learned dynamics model or the upstream supply of real-world action data used to train it.
Manufacturing scale-up
In the week ending May 2, 2026, five vendors reported manufacturing-scale milestones (Source: itcanthink.substack.com):
- Figure raised Figure 03 production to roughly 1 robot per hour (about 24 per day) and added stereo-camera 3D perception for navigation.
- 1x opened its Hayward factory, rated for 10,000 NEOs per year, and reported more than 10,000 NEO pre-orders in the first five days of launch.
- RobotEra (China) deployed L7 humanoids at China Post and other Chinese logistics hubs to align package barcodes, a task most retailers still pay humans for.
- Unitree launched a manipulation-focused line on its roughly $5,000 R1 platform with 19 degrees of freedom.
- Tutor Intelligence (Josh Gruenstein) unveiled "Data Factory 1," a 100-robot teleoperation facility scaling industrial-semi-humanoid training data.
Humanoid robots
Active humanoid programs include Figure AI (Figure paired with the Helix VLA model), Clone Robotics (biomimetic / musculoskeletal designs), Tesla Optimus, and Foxconn's humanoid program. Amazon was preparing to test humanoids for package delivery (June 2025). The underlying robotics technology in some humanoids has reportedly been Chinese-sourced (The Information, June 5, 2025).
Foundation models for robotics
Physical Intelligence describes its work as a "GPT for physical intelligence" with embodiment-flexible models. Helix is Figure's VLA model.
Google DeepMind introduced Gemini Robotics 2 on July 30, 2026 — a vision-language-action model, the Gemini Robotics ER 2 embodied-reasoning model, and Gemini Robotics On-Device 2. Controlling Apptronik's Apollo 2 humanoid with Inspire hands, it reported 68.4% success picking objects from a table, 76.3% from a shelf and 45.7% from the floor, with multi-finger tasks using SharpaWave hands ranging from 32% to 92%. DeepMind said On-Device 2 adapts to new bi-arm embodiments with typically fewer than 200 examples, and released ASIMOV-Agentic, a benchmark for agentic safety orchestration (Source: deepmind.google).
Autonomous vehicles
AV programs include Tesla Full Self-Driving, Waymo (robotaxis), and Zoox / Amazon. Mercedes-Benz and Nvidia announced planned robotaxis (January 29, 2026). GM / Cruise has had notable failures and withdrawals. A Waymo robotaxi struck a child in Santa Monica (TechCrunch, January 29, 2026), described as an accountability test. AV liability remains determined state-by-state, with the Cruise withdrawal cited as precedent.
Drones
Consumer drones are widely AI-enhanced. Research on "infinite flight" using laser wireless power was reported in early 2026 (Live Science, January 29, 2026). Walmart expanded drone delivery (WIRED, June 2025).
Other 2025–2026 developments
Additional developments in the period include questions about the Figure AI / BMW partnership (June 2025), the departure of the head of Tesla's Optimus program (Reuters, June 2025), and Physical Intelligence's ramp (TechCrunch, January 30, 2026).
Theoretical backdrop
The Industrial Explosion (Davidson/Hadshar, May 2025) frames humanoid robotics as the substrate of post-AGI physical-capability expansion. Its authors estimate initial doubling times of about one year, falling to days or weeks at capacity.
Regulatory and policy dimension
- ICRC position on autonomous weapons — international humanitarian law (IHL) limits on autonomous weapons systems; humanoid robots raise a dual-use concern.
- DOD Directive 3000.09 — Department of Defense directive on autonomous and semi-autonomous weapons.
- Lethal Autonomous Weapons (LAWS) — subject of UN discussions (Reuters, May 2025).
- UAS regulation — FAA rules on commercial drones.
- AV liability — determined state-by-state, with the Cruise case as precedent.
- FCC Covered List addition — market access. On July 28, 2026 the Federal Communications Commission added foreign-produced advanced robotic devices, defined as mobile robots including humanoids and quadrupeds, to its Covered List, barring new models from the equipment authorization required to import, market, or sell electronics in the United States.
The FCC action is the first instance of hardware-security policy reaching general-purpose robots rather than weapons or communications equipment, and it applies at the point of market entry rather than to use. Industry response split on whether it protects or constrains domestic development: Standard Bots Chief Executive Evan Beard called it "one of the strongest technology-security actions in modern U.S. history," while STM's Georg Stieler said it "could slow U.S. physical AI innovation by cutting startups and researchers off from future low-cost Chinese platforms before comparable Western alternatives exist," noting that China accounts for more than 90% of global magnet-rare-earth refining and sintered permanent-magnet production and has shipped 80% to 90% of humanoid robots to date (Source: therobotreport.com). China's Ministry of Commerce reported on July 29, 2026 that industrial robot exports rose 18.6% and 3D printer exports 109.3% year over year in the first half of 2026 (Source: theaiinsider.tech).
Physical Intelligence's π0.5 (2025) is a leading result on open-world generalization, which the paper calls "one of the biggest open problems in physical intelligence." It uses co-training on heterogeneous sources — multiple robots, high-level semantic subtask prediction, verbal instructions, and web data — with training examples that mix levels of abstraction, "combin[ing] image observations, language commands, object detections, semantic subtask prediction, and low-level actions." The reported result is a mobile manipulator cleaning kitchens and bedrooms "in new homes that were not present in the training data, performing complex multi-stage behaviors with durations of 10 to 15 minutes" — against a prior VLA baseline the paper characterizes as "typically less than a minute in length and often with relatively low success rates."
Senators Mike Rounds (R-SD) and John Hickenlooper (D-CO) introduced legislation, reported August 12, 2026, that would require the Commerce Department to develop voluntary standards for robotics use cases and to improve data sharing with industry on artificial intelligence models that "power advanced robotics," framed around competition with China. The retrievable text carries the bill title only as far as "The Robotics Supply Chain" and gives no bill number (Source: insideaipolicy.com).
Relationships
- supports: The Industrial Explosion (Davidson, Hadshar).
- depends-on: Multimodality, Reasoning Models and Chain-of-Thought.
- related: Figure AI, Physical Intelligence, Clone Robotics, Helix (Figure AI Vision-Language-Action model), Autonomous Weapons, ICRC Position on Autonomous Weapon Systems (source summary), DoD Directive 3000.09 — Autonomy in Weapon Systems (source summary), US-China AI Competition: Different Races, Different Metrics (defense robotics context).