# Robotics Foundation Model Industry Report

**Date:** March 2026 Echo Labs

---

## 1. Industry Overview

The robotics foundation model industry is experiencing explosive growth, with startups developing AI "brains" that can control robots across diverse embodiments and tasks. Rather than programming robots for specific functions, foundation models learn general-purpose physical intelligence — enabling a single AI to operate humanoids, robotic arms, quadrupeds, and mobile platforms.

### Market Size & Funding

- **$10.3B** in robotics startup funding in 2025, up **36%** from $7.54B in 2024
- Robotics funding grew **~2.64x** while the broader startup market grew only ~1.11x
- Q2 2025 alone saw ~$8.8B in deal value, with triple-digit quarterly increases
- The robot training data market is being called the **"gold rush of 2026"**, with the synthetic data generation market projected to grow at a 46.3% CAGR from 2025 to 2035

**Sources:**
- [Crunchbase: Robotics Funding Crests Higher](https://news.crunchbase.com/robotics/ai-funding-high-figure-raise-data/)
- [Crunchbase: Robotics Startup Funding Rises](https://news.crunchbase.com/robotics/startup-funding-rises-h1-2025-ai-apptronik-data/)
- [Robotics & Automation News: VC in Robotics](https://roboticsandautomationnews.com/2025/12/21/venture-capital-and-private-equity-in-robotics-where-is-the-smart-money-going/97794/)

---

## 2. Major Players

### 2.1 Skild AI — The Commercial Leader

| Detail | Value |
|---|---|
| **HQ** | Pittsburgh / San Francisco |
| **Founded by** | Carnegie Mellon robotics researchers |
| **Valuation** | $14B (Jan 2026) |
| **Total Raised** | ~$1.5B+ (Series C: $1.4B led by SoftBank) |
| **Revenue** | ~$30M (grew from $0 in months, 2025) |
| **Stage** | Commercially deployed |

**What they do:** Build the "Skild Brain" — a unified, omni-bodied foundation model that can control any robot type (humanoids, quadrupeds, arms, mobile manipulators) without retraining per embodiment.

**Technology & Data Strategy:**
- Pre-trains on **internet-scale human videos** (treating humans as "biological robots" to extract manipulation affordances)
- Generates **synthetic data** via NVIDIA Isaac Sim/Lab and Cosmos
- Fine-tunes with **< 1 hour of real robot data** per new skill
- Claims **1,000x more data points** than competing models
- Does **not purchase** teleop or synthetic data from third parties

**Revenue Model (software-only, no hardware):**
- Foundation model licensing via cloud APIs
- Specialized software modules per vertical
- Managed "AI factory" (private cloud training/inference)

**Key Customers & Partnerships:**
- **Foxconn + NVIDIA** — Deploying on assembly lines building NVIDIA Blackwell GPU servers (Houston, TX)
- **ABB Robotics** — Industrial deployment partnership
- **Universal Robots & MiR** (Teradyne) — OEM integration
- **HPE** — Infrastructure for local AI training/inference
- **Strategic investors/customers:** Samsung, LG, Schneider Electric, CommonSpirit Health, Salesforce

**Competitive Moat:** Deployment data flywheel — more robots deployed generates more real-world data, which improves the model for all customers.

**Sources:**
- [Skild AI Series C Announcement](https://www.skild.ai/blogs/series-c)
- [Skild AI + Foxconn/NVIDIA Factory Deal](https://technical.ly/entrepreneurship/pittsburgh-skild-ai-nvidia-foxconn-robotics-deployment/)
- [The Reindustrial Revolution (Skild blog)](https://www.skild.ai/blogs/reindustrial-revolution)
- [Learning by Watching Human Videos (Skild blog)](https://www.skild.ai/blogs/learning-by-watching)
- [Building the General-Purpose Robotic Brain (Skild blog)](https://www.skild.ai/blogs/building-the-general-purpose-robotic-brain)
- [Skild AI (NVIDIA case study)](https://www.nvidia.com/en-us/case-studies/skild-ai/)

---

### 2.2 Physical Intelligence (Pi) — The Data-Hungry Challenger

| Detail | Value |
|---|---|
| **Valuation** | Undisclosed (est. multi-billion) |
| **Total Raised** | ~$1B+ (Series B: $600M) |
| **Investors** | Thrive Capital, OpenAI, Lux Capital, Khosla Ventures, Sequoia |
| **Stage** | R&D + early deployments |

**What they do:** Build Pi-0, a generalist robot policy trained on large-scale multi-task, multi-robot data. Pursuing an open-source base model strategy.

**Data Strategy:**
- **Heavy buyer of data** — raised $600M to "collect more data and make strategic partnerships"
- Collects own teleoperation data at scale
- Uses simulation for synthetic data
- Open-source approach could commoditize the foundation layer

**Sources:**
- [Physical Intelligence raises $600M (Robot Report)](https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/)

---

### 2.3 World Labs (Fei-Fei Li) — The Spatial Intelligence Play

| Detail | Value |
|---|---|
| **Founded by** | Fei-Fei Li (Stanford, ImageNet creator) |
| **Total Raised** | ~$1.23B ($1B in Feb 2026) |
| **Investors** | AMD, Autodesk, NVIDIA, Fidelity, Emerson Collective, Sea |
| **Stage** | First product launched (Marble) |

**What they do:** Build "world models" — AI that understands 3D space, physics, collisions, and dynamics. Not directly controlling robots yet, but building the spatial intelligence layer that robotics depends on.

**Product:** Marble — generates editable 3D environments from text, photos, video, or panoramas. Available via freemium and paid tiers.

**Roadmap (3 phases):**
1. 3D space + time understanding (current)
2. Augmented reality
3. **Robotics** (future)

**vs. Skild:** World Labs is more foundational/upstream. If they succeed, they could become the spatial intelligence layer that robot brain companies build on — or compete with them directly.

**Sources:**
- [World Labs Raises $1B (The AI Insider)](https://theaiinsider.tech/2026/02/19/fei-fei-lis-world-labs-raises-1b-in-fresh-funding-to-advance-development-of-world-models/)
- [Spatial Intelligence is the Next Frontier (Fast Company)](https://www.fastcompany.com/91503667/world-labs-most-innovative-companies-2026)

---

### 2.4 AMI Labs (Yann LeCun) — The Contrarian Bet

| Detail | Value |
|---|---|
| **Founded by** | Yann LeCun (left Meta Nov 2025) |
| **Funding** | $1.03B seed (March 2026) — largest European seed ever |
| **Valuation** | $3.5B (pre-money) |
| **Stage** | Pre-product research |

**What they do:** Build world models based on **JEPA** (Joint Embedding Predictive Architecture) — a fundamentally different approach from LLMs. Instead of predicting tokens, JEPA learns abstract world representations from observation, like a baby learning gravity.

**Core thesis:** LLMs are fundamentally wrong for understanding the physical world. Token prediction cannot capture physical common sense.

**Data Strategy:**
- Trains on video, audio, sensor data, robot arm positions, lidar — **not text**
- Before leaving Meta, LeCun's team published **V-JEPA 2**: demonstrated zero-shot robot control in unseen environments

**First target markets:** Industrial robotics, healthcare, scientific research.

**vs. Skild:** A deeper architectural bet. If JEPA works, it could leapfrog the transformer-based approaches that Skild and most others use. But AMI Labs is years away from commercial deployment.

**Sources:**
- [Yann LeCun's AMI Labs Raises $1.03B (Latent Space)](https://www.latent.space/p/ainews-yann-lecuns-ami-labs-launches)
- [LeCun Raises $1B (Symplexia Labs)](https://news.symplexia.com/2026/03/technology-innovation/technology/yann-lecun-raises-1-billion-to-build-ai-that-understands-the-physical-world/)
- [World Models Race 2026 (Introl)](https://introl.com/blog/world-models-race-agi-2026)

---

### 2.5 Other Notable Competitors

| Company | Type | Valuation/Funding | Approach |
|---|---|---|---|
| **Figure AI** | Full-stack (hardware + brain) | ~$39B | Complete humanoid robots with integrated AI |
| **1X Technologies** | Full-stack | Raising at ~$10B | Humanoid robots; self-generates teleop data via "Expert Mode" |
| **Sanctuary AI** | Full-stack | — | Human-like robots for physical tasks |
| **Boston Dynamics** | Full-stack (Hyundai-owned) | — | Decades of integrated hardware-software experience |
| **Covariant** | Foundation model (acquired by Amazon) | — | RFM-1, 8B params; 6+ years proprietary warehouse data; 99% pick accuracy |
| **Genesis AI** | Foundation model | $105M seed (2025) | Synthetic physics engines; claims 3x faster training |
| **Rhoda AI** | Foundation model | $450M Series A (2026) | FutureVision model; motion priors needing only ~10hrs teleop |
| **NVIDIA (Project GR00T)** | Platform + model | Internal | Foundation model + tools for humanoids; both partner and competitor to all |
| **Dyna Robotics** | Foundation model | $120M Series A | — |
| **Archetype AI** | Physical agents | $35M | — |

---

## 3. The Data Landscape

Data is the critical bottleneck — and the biggest differentiator — in the robotics foundation model race. Three types of training data dominate:

### 3.1 Teleoperation Data

Humans physically control robots (via VR, joysticks, or exoskeletons) and every movement is recorded as training data. **High quality, but expensive and hard to scale.**

**Major Teleop Data Suppliers:**

| Company | Funding | What They Sell |
|---|---|---|
| **Scale AI** | $16.4B valuation | End-to-end "Physical AI Data Engine" — real-world teleop + synthetic data + labeling. The dominant player |
| **Cortex AI** (YC) | Early stage | Real-world workplace robot trajectories and egocentric human data. Runs a marketplace where workplaces get paid to host data collection |
| **Sensei** (YC) | Early stage | Robotic training data at scale via teleop-as-a-service |
| **micro1** | $35M raised, $500M val | Human interaction videos turned into robotics training data via global collector network |
| **PrismaX** | Early stage | Crowdsourced teleop via VR; pays operators $50/hr; sells logged datasets |
| **General Intuition** | $134M raised | Gameplay video data applicable to robotic systems |

**Sources:**
- [Scale AI Physical AI Data Engine](https://scale.com/physical-ai)
- [Cortex AI](https://cortexrobot.ai/)
- [Sensei (YC)](https://www.ycombinator.com/companies/sensei)
- [micro1 Robotics Data Engine](https://www.micro1.ai/data-engine/robotics)

### 3.2 Synthetic Data

Generated via physics simulations. **Infinitely scalable** (just add GPUs), but sim-to-real transfer remains a challenge.

**Major Synthetic Data Providers:**

| Company | What They Provide |
|---|---|
| **NVIDIA (Cosmos, Isaac Sim, Omniverse)** | The dominant simulation/synthetic data platform. Used by nearly everyone |
| **Datagen** (acquired by NVIDIA) | Synthetic visual data for AI |
| **Parallel Domain** | Synthetic environments for robotics and autonomous vehicles |
| **Synthesis AI** | Synthetic data for vision models |

NVIDIA also released the **Physical AI Data Factory Blueprint** — an open reference architecture for generating, augmenting, and evaluating training data at scale. Early adopters include Skild AI, Uber, Teradyne Robotics, FieldAI, and others.

**Sources:**
- [NVIDIA Physical AI Data Factory Blueprint](https://nvidianews.nvidia.com/news/nvidia-announces-open-physical-ai-data-factory-blueprint-to-accelerate-robotics-vision-ai-agents-and-autonomous-vehicle-development)
- [Scaling Physical AI with Synthetic Data (NVIDIA blog)](https://blogs.nvidia.com/blog/scaling-physical-ai-omniverse/)

### 3.3 Internet Video Data

Publicly available human videos scraped from the web (e.g., YouTube). **Free and massive scale**, but requires sophisticated techniques to extract actionable robotics data.

Skild AI is the primary proponent of this approach, treating humans as "biological robots" and extracting manipulation affordances from billions of online videos.

### 3.4 Who Buys What — Summary

| Company | Buys Teleop? | Buys Synthetic? | Self-Generates? | Primary Strategy |
|---|---|---|---|---|
| **Skild AI** | No | No (generates own via NVIDIA tools) | Yes — internet videos + simulation + deployments | Avoid expensive data entirely |
| **Physical Intelligence** | Yes (heavy buyer + own collection) | Uses NVIDIA tools | Yes | Buy + collect at scale |
| **Figure AI** | Likely (own + purchased) | NVIDIA tools | Yes — proprietary humanoid deployments | Vertical data advantage |
| **1X Technologies** | Self-generates ("Expert Mode") | Yes | Yes — deployment data loop | Clever self-collection |
| **World Labs** | Not yet (robotics is phase 3) | Likely building own | 3D/spatial data | Pre-robotics stage |
| **AMI Labs** | TBD (very early) | TBD | Video, audio, sensor data | Architecture-first |
| **NVIDIA (GR00T)** | Partners provide | Owns the platform | Internal R&D | Sells picks and shovels |

---

## 4. Three Strategic Bets

The industry is splitting into three competing theses on how to build physical AI:

### Bet 1: "Buy and collect real-world data at scale"
**Players:** Physical Intelligence, Figure AI, Scale AI ecosystem
**Thesis:** Real-world data is irreplaceable. Simulation and video can't capture the full complexity of physical interaction. Spend heavily to collect and purchase teleop data.
**Risk:** Expensive and slow. If synthetic/video approaches work, this becomes a cost disadvantage.

### Bet 2: "Generate everything synthetically + internet video"
**Players:** Skild AI, Genesis AI
**Thesis:** Teleoperation doesn't scale. Internet videos + physics simulation provide 1,000x more data at a fraction of the cost. Fine-tune with minimal real-world data.
**Risk:** Sim-to-real gap. Synthetic data may not capture edge cases that matter in deployment.

### Bet 3: "New architecture entirely"
**Players:** AMI Labs (JEPA), World Labs (spatial intelligence)
**Thesis:** The current transformer/token-prediction paradigm is fundamentally wrong for physical AI. A new architecture (JEPA, world models) is needed to achieve true physical common sense.
**Risk:** Unproven at scale. Years away from commercial deployment. The transformer crowd may iterate faster.

---

## 5. Competitive Comparison Matrix

| | Skild AI | Physical Intelligence | World Labs | AMI Labs | Figure AI | 1X |
|---|---|---|---|---|---|---|
| **Valuation** | $14B | Multi-B (est.) | ~$1.25B+ | $3.5B | ~$39B | ~$10B (est.) |
| **Stage** | Deployed, revenue | Early deployment | First product | Pre-product | Deployed | Deployed |
| **Revenue** | ~$30M | Undisclosed | Early | None | Undisclosed | Undisclosed |
| **Approach** | Software-only brain | Software-only brain | World models | JEPA world models | Full-stack humanoid | Full-stack humanoid |
| **Data strategy** | Video + sim | Buy + collect teleop | Own 3D data | Video/sensor/audio | Own teleop + bought | Self-generated teleop |
| **Key advantage** | Deployment flywheel | Data scale | Fei-Fei + spatial IP | JEPA architecture | Owns full stack | Expert Mode data loop |
| **Key risk** | Competition | Cost of data | Robotics is phase 3 | Unproven, years away | Hardware margins | Scaling manufacturing |

---

## 6. Key Takeaways

1. **NVIDIA is the universal winner.** Regardless of which foundation model company wins, nearly all of them depend on NVIDIA's simulation stack (Isaac Sim, Cosmos, Omniverse) and GPUs. NVIDIA is both infrastructure provider and competitor (via GR00T).

2. **Data is the moat.** The companies with the best data flywheels — whether through deployment (Skild, 1X), purchasing (Physical Intelligence), or novel collection (internet video) — will likely win.

3. **Software-only vs. full-stack is the key strategic divide.** Skild and Physical Intelligence sell the brain to any hardware maker. Figure and 1X build the whole robot. Both approaches have merits; the market may support both.

4. **The "anti-LLM" bet is the wildcard.** LeCun's AMI Labs is betting the entire transformer paradigm is wrong for physical AI. If JEPA works, it reshuffles the entire competitive landscape. If not, AMI Labs becomes an expensive research project.

5. **Commercialization is accelerating.** Skild went from $0 to ~$30M revenue in months. The industry is transitioning from pure R&D to real deployment — security, warehouses, manufacturing, and data centers are the beachhead markets before consumer homes.

---

*Report compiled March 2026. All funding figures, valuations, and revenue numbers are based on publicly available information and may not reflect the most current data.*