Waddle Labs, backed by Y Combinator and staffed by just two founders, launched an API that lets developers control industrial robots through typed instructions, generating executable code in roughly 20 minutes. The San Francisco outfit calls itself "Claude Code for robotics," a pitch from CEO Hanming Ye that lands at a moment when the robotics industry is scrambling to connect foundation models to factory floors populated by 4.66 million industrial robots that still require weeks of specialized integration work per installation.
The timing reflects broader momentum. Industrial robotics hit 542,000 new installations in 2024—2 percent lower than the 2022 peak, according to the International Federation of Robotics. Global operational stock climbed 9 percent year-over-year to reach 4.66 million units by year-end 2024, data that the Stanford AI Index later cited. China claimed over half of new placements. Europe saw installations decline 8 percent to 85,000 units.
But programming those robots remains glacial. Traditional industrial automation demands months of system integration, specialized motion planning, and per-task code written by roboticists fluent in kinematics, trajectory libraries, and vendor-specific APIs. Waddle's hypothesis: large language model agents can marry the generalization of vision-language-action models with the speed and reliability of classical control methods. "The agent decomposes long tasks into stages, verifies intermediate outcomes, and re-plans on failure," the company wrote in a research post published mid-2026.
Waddle's architecture features three layers. A hardware abstraction SDK handles safety envelopes. An intermediate representation layer tackles inverse kinematics and trajectory planning, making the system hardware-agnostic—documentation notes that porting to a new arm or camera requires minimal adapter work. A closed-source agent harness manages model routing and code execution. The design mirrors a trend in 2026 academic and industrial labs toward "code-as-policy" frameworks, though Waddle's founders declined to disclose customer names, revenue figures, or funding details beyond their Y Combinator batch.
What Changed to Make This Plausible
Three technical advances converged recently. Foundation models scaled. NVIDIA released its GR00T N1 open model for humanoids in March 2025, iterating through versions N1.5 and N1.6 by late 2025, with later posts describing GR00T 1.7 workflows spanning data collection, simulation, evaluation, and on-robot deployment using Jetson Thor compute. Google DeepMind followed with Gemini Robotics 2 in August 2026, demonstrating whole-body dexterous manipulation across multiple robot types while emphasizing that many applications require on-device, low-latency operation for reliability.
Dataset scale reached critical mass. The Open X-Embodiment project published over 1 million real robot trajectories across 22 embodiments and 527 skills at ICRA 2024; subsequent work like OXE-AugE added synthetic augmentations. Covariant deployed its RFM-1 foundation model to what it described as "dozens of customers" across more than 15 countries starting March 2024, running bin-picking, induction, and depalletizing tasks in live warehouses for clients including Radial and the bpost group. OpenVLA and Octo, both released in 2024, became widely used baselines in subsequent research.
Code-generation agents proved they could write, debug, and refine robotic control policies autonomously. CaP-X, a framework published in early 2026, reported meaningful zero-shot success with training-free coding agents and reinforcement-learned variations. ASPIRE, released mid-2026, introduced a self-improving agent that accumulates reusable skills into a persistent library, achieving gains on long-horizon LIBERO benchmarks. Anthropic published "Claude plays robotics" demonstrations around the same time, showing its models operating real robots via code and visual interfaces.
McKinsey estimated in a November 2025 report that existing technology could potentially automate 57 percent of current U.S. work hours, with manufacturing at 31 percent. The firm emphasized scaling "a focused portfolio of 5–12 use cases by 2030," often centered on shop-floor automation and robotics control systems. Labor shortages and product-mix variability push demand beyond rigid automation. AI agents promise faster re-tasking without months of system integration, according to reports from BCG and the World Economic Forum published between late 2025 and mid-2026.
A Crowded Field
Waddle Labs enters a space already populated by well-funded competitors. Alphabet's Intrinsic operates Flowstate, a production-grade industrial robotics development environment with integrations to NVIDIA Omniverse, Isaac, and ROS 2. The company demonstrated machine-tending workflows at IMTS in September 2024 and announced a joint venture with Foxconn in November 2025 to build what it called an "AI factory of the future."
Covariant's RFM-1, launched March 2024, runs in warehouses today. The company claims "dozens of customers" in more than 15 countries, with public case studies from logistics operators performing bin-picking and depalletizing tasks. Figure AI and BMW announced trials of humanoid robots at BMW's Spartanburg plant starting in 2025; Figure posted updates noting its F-03 model at BMW as recently as mid-2026.
NVIDIA frames the shift as "physical AI." CEO Jensen Huang has described factories as "robots orchestrating robots" in a three-computer stack: training, simulation, and edge inference. The company's Jetson Thor modules, detailed in datasheets, pair with TensorRT Edge-LLM to support on-device inference of quantized foundation models at sub-100-millisecond latencies. Agility Robotics deployed its Digit humanoid under a multi-year robotics-as-a-service agreement with GXO starting in 2024, and Amazon tested units around the same time, though neither deployment centers on LLM-direct control.
Waddle founders Hanming Ye and Yiding Song both list Harvard affiliations on LinkedIn. Ye's publications include "Steering Diffusion Policies with Value-Guided Denoising," while Song worked on multimodal astrophysics projects including PAPERCLIP. The startup completed Y Combinator's Summer 2026 batch with Ankit Gupta as primary partner.
The Safety Problem
Industrial robot standards expect deterministic risk-reduction measures. ISO 10218-1:2025 and 10218-2:2025, which came into force April 2025 and absorbed collaborative-robot requirements from ISO/TS 15066:2016, don't mesh naturally with probabilistic models. LLMs and vision-language-action models are, by design, non-deterministic.
A survey published in 2026 titled "Robotic Foundation Models for Industrial Control: Readiness Assessment" evaluated 324 models across 149 criteria, arguing that progress depends on safety validation, real-time feasibility, robust perception, and auditable integration stacks. Security poses additional risks. Research from the past two years demonstrated prompt-injection, jailbreak, and adversarial perception attacks against VLM- and VLA-controlled systems, highlighting the need for authenticated I/O, safety supervisors, and input guardrails.
The EU AI Act, with consolidated text updated in mid-2026, classifies AI used as safety components in products covered by sectoral harmonization as high-risk, requiring conformity assessment. NIST's Generative AI Profile, published July 2024, lays out risk-management expectations for probabilistic models in critical applications.
Latency and network isolation matter on factory floors. Many industrial sites require on-premises or on-device operation due to network downtime risk and IT security policies—Google DeepMind's Gemini Robotics 2 blog post explicitly noted the need for on-device inference. ISA/IEC 62443 cybersecurity standards form the baseline for industrial control networks, relevant when connecting LLM agent tools to programmable logic controllers and motion systems.
OSHA's Technical Manual Section IV, Chapter 4 provides U.S. baseline guidance for robot system safety inspections and references RIA and ISO documentation. ANSI/A3 R15.08 Part 1 covers industrial mobile robots; Part 2 addresses systems and applications. Recent work explores LLM-assisted PLC programming with vendor overlays and validation layers to bridge probabilistic outputs to deterministic IEC 61131-3 logic, peer-reviewed studies show.
An Uncertain Horizon
Market projections vary wildly. Goldman Sachs Research estimated a $38 billion humanoid robot market by 2035 in an early 2024 baseline forecast, with later summaries suggesting higher unit volumes: 75,000 units in 2026, 890,000 in 2030, and roughly 6.5 million per year by 2035. UBS wrote in mid-2025 that 63 percent of 2030 humanoid use cases would be industrial, with shipment growth tied to AI and compute advances plus regulatory clarity.
Waddle's hardware-agnostic intermediate representation and 20-minute time-to-first-policy target a specific pain point: the integration bottleneck that has kept foundation models out of most factories despite billions in research investment. Independent third-party benchmarks and industrial pilot references for Waddle were not yet publicly available at last check. The company declined to disclose customer names, revenue, or funding details beyond its Y Combinator batch.
The architecture Waddle has adopted mirrors choices in CaP-X, ASPIRE, and recent NVIDIA research—code-as-policy with multi-agent orchestration, iterative skill refinement, and a shared intermediate representation. Differences will hinge on reliability under plant conditions, integration with safety PLCs, and the test-evaluate-verify-validate rigor required by ISO 10218:2025 and EU AI Act high-risk classifications.
IFR President Takayuki Ito said in a 2025 report that 2024 "was the second highest annual installation count... only 2 percent lower than the all-time high two years ago." The 542,000 installations that year, and the 4.66 million robots now in operation, represent a hardware base waiting for software that can reprogram itself quickly. Whether prompt-driven agents deliver on that promise at industrial scale remains an open engineering question. But the race to answer it is already crowded, and Waddle—population two—has joined.
