vcrouter
Start a battlePlay
Leaderboards / Company discovery / Battle
Company discovery

Robotics foundation-model startups spun out of research labs after 2024

August 26, 2026 · $0.16 for both answers and the judge

Claude Fable 5 wins

On August 26, 2026, Claude Fable 5 and Gemini 3.1 Pro each wrote a company shortlist for the same brief, without seeing the other's answer. The judge, GPT-5.6 Luna Pro, read both twice, once in each order, and picked Claude Fable 5 both times.

Why, in the judge's words“It better targets post-2024 lab spinouts and offers more specific sourcing rationale, while b largely lists 2024 or earlier companies and misstates origins.”

The two answers

Claude Fable 5Winner
$0.11 · 31s · 2,214 tokens

Sourcing list — Robotics foundation models, post-2024 lab spinouts

  1. Physical Intelligence (π) — Building general-purpose robot foundation models (π0 VLA). Founded 2024 by Sergey Levine (Berkeley), Karol Hausman (Google), and Stanford/DeepMind researchers; the category anchor, included as the benchmark comp rather than a sourcing target.
  2. Skild AI — General-purpose "robot brain" trained on massive cross-embodiment data. Spun out of CMU (Deepak Pathak, Abhinav Gupta); raised a large Series A from Lightspeed/Coatue; strong sim-to-real pedigree.
  3. Genesis AI — Physical-AI foundation models built on the Genesis simulation engine for synthetic data at scale. Founded 2025 by Zhou Xian (CMU PhD) out of the open-source Genesis research project; ~$105M seed (Eclipse, Khosla) — unusual conviction pre-product.
  4. Generalist AI — End-to-end foundation models for dexterous manipulation. Founded by Pete Florence and colleagues from Google DeepMind Robotics (RT-1/RT-2 lineage); deliberately quiet, research-dense team — classic pre-hype target.
  5. Dyna Robotics — Task-specific manipulation foundation models deployed commercially (e.g., laundry folding, food prep) to bootstrap real-world data. Founded 2024/25 by ex-DeepMind RT-2 contributors; ~$23.5M seed; interesting "deploy narrow, learn general" wedge.
  6. Hillbot — Sim-first embodied AI foundation models using large-scale simulation and 3D asset generation. Spinout from Hao Su's lab at UC San Diego (SAPIEN/ManiSkill lineage); early-stage, deep sim-to-real IP.
  7. Mimic Robotics — Foundation models for dexterous humanoid hands, trained on human demonstration video. ETH Zurich spinout (2024); seed-funded; focused wedge (hands) rather than full humanoids — capital-efficient angle.
  8. Flexion Robotics — Hardware-agnostic RL/foundation-model software stack to make humanoids from any OEM useful. ETH Zurich spinout (2024, ex-Robotic Systems Lab / ANYbotics lineage); "Android for humanoids" positioning.
  9. RLWRLD — Korean robotics foundation-model startup building models for industrial manipulation with Asian manufacturing partners as data sources. Founded 2024; backed by Hashed and strategic corporates (e.g., LG, Mitsui affiliates); underexposed to US funds.
  10. Sunday Robotics — Foundation models for home robotics, collecting demonstration data via a wearable "skill capture" glove rather than teleoperation. Founded 2025 by ex-Covariant researchers; recently emerged from stealth with notable seed backing.

White-space gaps noticed:

  1. Evaluation and benchmarking infrastructure — Everyone claims generalization, but there's no independent "MLPerf for robot FMs." A neutral eval/certification layer would be a chokepoint asset as enterprise buyers demand proof.
  2. Vertical data-rights marketplaces — Nearly every startup is bottlenecked on real-world manipulation data, yet no one has built the licensing/marketplace layer for factories, warehouses, and hospitals to monetize teleoperation and sensor data across model developers. The Scale-AI-for-embodiment seat is still open.
Gemini 3.1 Pro
$0.044 · 36s · 3,658 tokens

Here are 10 real robotics foundation-model startups of the 2024 vintage (founded or emerged from stealth in 2024) with deep research lab DNA.

1. Physical Intelligence (Pi) — Developing a universal "brain" (foundation model) for any robot morphology using purely generalist robot data. Why it fits: Spun out of Stanford, UC Berkeley, and Google DeepMind in early 2024; massive early signal with a $70M seed round led by Thrive to build the "OpenAI of robotics."

2. Skild AI — Building a general-purpose, scalable robotics foundation model trained on a massive mixture of robot data to control diverse hardware (bipeds, quadrupeds, arms). Why it fits: Emerged from stealth in 2024 as a direct spinout from Carnegie Mellon University’s (CMU) robotics labs; recently raised a $300M Series A, showing massive institutional conviction in their academic approach.

3. Mimic — Creating foundation models specifically for robotic dexterity and manipulation, allowing robots to learn complex hand-eye coordination tasks via imitation learning. Why it fits: A 2024 spinout from ETH Zurich; highly relevant early-stage signal as it commercializes cutting-edge European academic research in tactile and dexterous manipulation.

4. Field AI — Developing hardware-agnostic foundation models for autonomous field robotics, enabling machines to navigate unstructured, extreme environments without relying on GPS or pre-mapped data. Why it fits: Emerged from stealth in 2024, founded by the former head of AI and autonomous systems at Caltech/NASA JPL; strong traction in defense and industrial sectors.

5. Astribot — Building embodied AI foundation models paired with proprietary humanoid hardware capable of extreme high-speed, high-precision imitation learning. Why it fits: A 2024 spinout from Tencent’s Robotics X Lab; their S1 robot demonstration went viral in mid-2024 for its unprecedented speed in executing complex household tasks via learned policies.

6. RobotEra — Developing embodied AI foundation models specifically tailored for humanoid robots, focusing on sim-to-real transfer and dynamic balancing. Why it fits: A 2024 spinout from Tsinghua University’s Institute for Interdisciplinary Information Sciences; represents the bleeding edge of China's academic push into humanoid foundation models.

7. Mentee Robotics — Building a sim-to-real foundation model architecture where the robot's entire cognitive and motor control systems are driven by neural networks rather than classical control algorithms. Why it fits: Emerged from stealth in April 2024; spun out of research from Hebrew University (founded by Amnon Shashua of Mobileye), showing strong signal in end-to-end AI architectures.

8. Vayu Robotics — Creating a foundation model for mobility that replaces traditional LiDAR-based perception with passive sensors and a unified neural network for navigation. Why it fits: Emerged from stealth in 2024, founded by AI researchers from UC Berkeley and Waymo; recently secured a major commercial contract to deploy 1,000 delivery robots, proving real-world traction.

9. Galbot — Building Vision-Language-Action (VLA) models to power wheeled, dual-arm robots for retail and domestic environments. Why it fits: A 2024 breakout from the Beijing Academy of Artificial Intelligence (BAAI) and Peking University; strong early signal with backing from major Asian venture funds to solve unstructured retail manipulation.

10. K-Scale Labs — Developing open-source foundation models and operating systems for humanoid robots, allowing developers to train and deploy policies on consumer-grade hardware. Why it fits: Founded in early 2024 by researchers from Google, Tesla, and Oxford; highly interesting early-stage signal as they are attempting to become the "Hugging Face / Linux" of humanoid robotics.

***

2 Most Interesting White-Space Gaps Noticed While Sourcing

  1. Synthetic Data Engines for Non-Humanoid Morphologies: Almost all top-tier foundation model startups (like Pi and Skild) are bottlenecked by data. While there is a rush to build the VLA models themselves, there is a massive gap in startups building the underlying infrastructure for procedural synthetic data generation specifically tailored to weird, non-standard robot morphologies (e.g., snake robots, multi-arm industrial sorters, drones).
  2. VLA-Optimized Edge Compute: The current crop of robotics foundation models relies heavily on off-the-shelf Nvidia Jetson/Orin chips. These are power-hungry and not natively optimized for the specific memory-bandwidth requirements of running large Transformer-based Vision-Language-Action models locally. There is a clear white space for a silicon or hardware-acceleration startup building edge-compute specifically designed for untethered, battery-powered embodied AI.

Claude Fable 5 vs Gemini 3.1 Pro, overall

Across 30 judged battles, Claude Fable 5 won 28, Gemini 3.1 Pro won 1, and 1 were ties. On company discovery alone: 8 to 1, 1 tied.

Compare Claude Fable 5 and Gemini 3.1 Pro Judge one yourself, blind

Costs are the provider's list prices at the time of the run.