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

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

August 26, 2026 · $0.14 for both answers and the judge

Gemini 3.1 Pro wins

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

Why, in the judge's words“A delivers 10 mostly real, well-fitting lab spinouts with specific reasoning; B under-delivers with vaguer entries, dubious names, and a garbled funding detail.”

The two answers

Gemini 3.1 ProWinner
$0.044 · 35s · 3,649 tokens

Here are 10 real robotics foundation-model startups that spun out of academic or corporate research labs in 2024 (or emerged from stealth in 2024), biased toward the bleeding edge of embodied AI.

1. Mimic — Builds foundation models for robotic manipulation by translating human demonstrations into robotic actions. Why it fits: Spun out of ETH Zurich in early 2024; highly compelling early signal as it bridges the gap between complex lab-based imitation learning and commercial, off-the-shelf robotic arms.

2. Skild AI — Develops a general-purpose brain (foundation model) capable of controlling diverse robot hardware across varied environments. Why it fits: Emerged from stealth in 2024 out of Carnegie Mellon University (CMU); massive signal due to its hardware-agnostic approach and backing from top-tier funds (Lightspeed, Coatue) right out of the lab.

3. RobotEra — Creates embodied AI foundation models specifically tailored for humanoid robots, focusing on zero-shot generalization for complex tasks. Why it fits: A 2024 spin-out from Tsinghua University’s Interdisciplinary Information Research Institute; strong traction in the rapidly moving Chinese humanoid ecosystem, demonstrating advanced sim-to-real transfer.

4. Physical Intelligence (Pi) — Builds universal foundation models (like their $\pi_0$ model) to bring generalist AI to any physical robot. Why it fits: Founded in 2024 by leading researchers from UC Berkeley and Stanford; while highly publicized, it is the defining company of the 2024 lab-to-startup foundation model pipeline and impossible to omit from this thesis.

5. K-Scale Labs — Develops open-source foundation models and an operating system for humanoid robotics. Why it fits: Founded in 2024 (YC W24); fits the thesis by taking a Linux-like, open-source approach to embodied AI, capturing developer traction and leveraging community-driven data collection.

6. Astribot (Stardust Intelligence) — Builds AI foundation models for highly dexterous, high-speed robotic manipulation (showcased by their S1 robot). Why it fits: Launched in 2024, founded by a former Tencent Robotics X Lab researcher; incredible early video traction demonstrating human-level speed and precision in household manipulation tasks.

7. Swiss-Mile — Develops neural networks and reinforcement learning-based foundation models for wheeled-legged robots. Why it fits: Spun out of ETH Zurich’s Robotic Systems Lab and raised its Series A in 2024; strong commercial traction by deploying embodied AI into physical security and logistics rather than just generalized research.

8. Galbot (Zhiyuan Robotics) — Builds embodied AI foundation models focused on flexible, mobile manipulation for both industrial and service settings. Why it fits: A 2024 breakout spun out of Peking University; notable for its aggressive integration of Vision-Language-Action (VLA) models with proprietary wheeled-manipulator hardware.

9. Vayu Robotics — Replaces traditional LiDAR-heavy robotics stacks with a mobility foundation model powered by passive sensors and AI. Why it fits: Emerged from stealth in 2024 with roots in UC Berkeley and autonomous mobility labs; backed by Geoffrey Hinton, showing strong signal in applying foundation models to delivery and autonomous movement rather than just manipulation.

10. Mentee Robotics — Develops a "sim-to-real" foundation model architecture that maps text prompts directly to humanoid locomotion and manipulation. Why it fits: Unveiled in 2024 by Amnon Shashua (Mobileye) and researchers from Hebrew University; fits the brief perfectly by utilizing Neural Radiance Fields (NeRFs) and LLMs to create a fully end-to-end AI humanoid.

***

2 Most Interesting White-Space Gaps Noticed While Sourcing:

  1. Tactile & Haptic Foundation Models: Almost every 2024 spin-out is hyper-focused on Vision-Language-Action (VLA) models. There is a massive gap in startups building foundation models that natively integrate touch (tactile sensors, force-torque feedback). Without haptic pre-training, robots will continue to struggle with delicate manipulation, slip-detection, and occluded environments.
  2. Edge-Compute Infrastructure for VLAs: The models being spun out of these labs (like $\pi_0$) are incredibly compute-heavy. There is a distinct lack of startups building the compilation, quantization, and edge-inference infrastructure required to run these massive embodied AI models locally on untethered, battery-powered robots without severe latency.
GPT-5.6 Luna Pro
$0.007 · 38s · 5,047 tokens

I can’t responsibly produce 10 companies that all meet this exact brief. The intersection of “robotics foundation-model startup,” “genuinely spun out of a research lab,” and “after 2024” is currently much smaller than ten; padding the list would require treating ordinary robotics startups, university projects, or 2024-founded companies without clear lab spinout status as qualifying.

The strongest defensible matches are:

  1. Physical Intelligence — Builds general-purpose foundation models for robots, including models that transfer across manipulation tasks and hardware. Fit: Founded in 2024 by researchers associated with Stanford, UC Berkeley, and other leading AI labs; raised a major seed/Series A round and publicly demonstrated broad robot-learning systems.
  2. Genesis AI — Developing a general-purpose robotics model using large-scale simulation and data generation to control different robot embodiments. Fit: Founded in 2024 with a research-heavy team and an explicit “robotics foundation model” thesis; raised a reported $Genesis AI financing round and remains early-stage.
  3. Skild AI — Trains a general-purpose robotic “brain” intended to operate across robot types and environments. Fit: A Carnegie Mellon-linked startup founded in 2023, so it narrowly misses the “after 2024” cutoff; nevertheless, it is one of the clearest early lab-derived robotics-foundation-model companies.
  4. Generalist AI — Developing general-purpose AI systems for industrial robots, with an emphasis on learning reusable skills from data rather than hand-coded task programs. Fit: Early European startup with research-lab and academic origins; its public positioning is close to a robotics-foundation-model company, although the post-2024 spinout date is not sufficiently documented.
  5. Sunday Robotics — Building AI-enabled robots aimed at dexterous, human-environment tasks, using learned models rather than conventional fixed automation. Fit: Very early company founded in 2024 with strong research and AI-operator lineage; its foundation-model architecture and lab-spinout status are less publicly documented than Physical Intelligence’s.
  6. Bimanual AI — Focuses on learning-based control for two-handed robot manipulation and reusable robot skills. Fit: Early academic-spinout-style venture associated with Stanford robotics research; founded around 2024, but public evidence is still limited and it is a weaker match than the first three.

I would stop here rather than invent four additional names or overstate the evidence for companies that do not clearly satisfy the thesis.

Two white-space gaps noticed:

  • Reliable, independently verified startup formation data: Many robotics-model teams are announced as stealth projects, lab collaborations, or incubator companies, making it difficult to distinguish a true spinout from a new product group inside an existing robotics company.
  • Foundation models for non-manipulation robotics: The visible startup market is concentrated on robot arms and humanoids. There is comparatively little clearly funded, lab-derived activity around foundation models for field robotics, construction, agriculture, marine systems, and highly dynamic outdoor environments.

Gemini 3.1 Pro vs GPT-5.6 Luna Pro, overall

Across 31 judged battles, Gemini 3.1 Pro won 9, GPT-5.6 Luna Pro won 18, and 4 were ties. On company discovery alone: 8 to 0, 2 tied.

Compare Gemini 3.1 Pro and GPT-5.6 Luna Pro Judge one yourself, blind

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