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

Here is a curated list of 10 robotics foundation-model startups that were founded, spun out, or unstealthed in 2023/2024 (as strictly post-2024 foundings are still largely in stealth). I have prioritized academic and corporate lab spinouts building general-purpose embodied AI, leaning toward lesser-known players where possible.

1. Mimic — Develops foundation models for robotic manipulation with a focus on collaborative and industrial arms. Why it fits: Spun out of ETH Zurich in 2024. Raised a $2.5M pre-seed; strong signal from the European academic robotics ecosystem, leveraging proprietary data collection methods rather than relying solely on open-source datasets.

2. Physical Intelligence (Pi) — Building a universal foundation model (π0) to bring general-purpose AI to any physical robot. Why it fits: Founded in 2024 as a super-spinout from UC Berkeley, Stanford, and Google DeepMind. Raised a massive $400M at a $2.4B valuation; the ultimate signal in the VLA (Vision-Language-Action) space.

3. Archetype AI — Building "Newton," a multimodal foundation model that parses raw sensor data (radar, IMU, cameras) to understand physical world physics and behavior. Why it fits: Unstealthed in 2024 (founded late 2023) by the former Google ATAP (Advanced Technology and Projects) team. Raised a $13M Seed; highly unique because it focuses on complex sensor fusion rather than just vision-to-action.

4. Skild AI — Developing a scalable, general-purpose brain for diverse robotic hardware, trained on a massive cross-embodiment dataset. Why it fits: Unstealthed in 2024 with a $300M Series A. Spun out of Carnegie Mellon University (CMU) by Deepak Pathak and Abhinav Gupta. Massive traction and academic signal in cross-embodiment generalization.

5. Galbot — Creating embodied AI foundation models specifically tailored for mobile manipulation in retail and domestic environments. Why it fits: Spun out of Tsinghua University and the Beijing Academy of Artificial Intelligence (BAAI). Raised a ~$97M Series A in 2024; represents the leading edge of China's aggressive push into commercializing lab-born VLA models.

6. K-Scale Labs — Building open-source foundation models and operating systems specifically for humanoid robots. Why it fits: Founded in 2024 by former Tesla Optimus and AI researchers. Backed by Y Combinator (W24); strong developer traction by open-sourcing their models and hardware designs to crowdsource real-world data.

7. RobotEra — Developing embodied AI foundation models integrated with proprietary humanoid hardware for general-purpose tasks. Why it fits: Spun out of Tsinghua University’s Institute for Interdisciplinary Information Sciences. Raised a ~$40M Series A in August 2024; strong academic pedigree and rapid iteration on walking/manipulation algorithms.

8. Sereact — Building transformer-based foundation models (PickGPT) that combine LLMs with computer vision for zero-shot robotic grasping and warehouse manipulation. Why it fits: Spun out of the University of Stuttgart. Raised a $5M Seed; highly pragmatic application of lab research, showing immediate commercial traction in logistics rather than waiting for AGI.

9. Astribot — Developing highly dexterous, AI-driven humanoid assistants powered by proprietary imitation-learning foundation models. Why it fits: Founded in 2024 by researchers out of Tencent Robotics X Lab. Recently showcased their S1 robot performing incredibly high-speed, human-level manipulation tasks; strong stealth-to-viral traction.

10. Prosper Robotics — Building foundation models for household robots capable of doing dishes, laundry, and cleaning via advanced mobile manipulation. Why it fits: Founded by Shariq Hashme (ex-OpenAI). Currently in early stages/stealth testing in London; strong founder-market fit combining OpenAI-style scaling laws with consumer robotics.

***

2 Most Interesting White-Space Gaps

1. Teleoperation & Data-Collection Infrastructure Every foundation model startup is bottlenecked by the same thing: high-quality, real-world physical data. While companies are building the "brains," there is a massive white space for startups building the "picks and shovels" of data collection—specifically, low-latency, haptic-feedback teleoperation rigs, standardized data-ingestion pipelines, and synthetic data generation engines tailored strictly for VLA models.

2. Edge-Compute Optimization for VLAs Running massive, multi-billion parameter Vision-Language-Action models requires immense compute, which drains the batteries of untethered robots and causes latency that ruins physical manipulation. There is a glaring gap for startups building specialized compilation tools, quantization frameworks, or novel silicon architectures specifically designed to run embodied AI foundation models locally on power-constrained robots.