Industry Map · May 8, 2026 · 7 min read
Who's Actually Building Physical AI
A plain-language map of the physical AI ecosystem: foundation models, data platforms, open-source infrastructure, and managed data ops.
Who's Actually Building Physical AI
Foundation models. Data platforms. Annotation tools. Open-source infrastructure. The physical AI ecosystem has multiple distinct layers - and most people conflate them. Here's a plain-language breakdown of who's doing what.
The physical AI industry is moving fast enough that it's hard to track who's who. Skild AI and Physical Intelligence are often mentioned in the same breath as Scale AI - but they're doing completely different things. Encord is sometimes called a "Scale AI competitor" but the comparison is imprecise. Hugging Face's LeRobot is open-source infrastructure, not a product company.
The confusion matters. If you're building a robot, training a policy, or deciding where to invest, knowing what layer of the stack each player operates in is the whole game. This article maps it out.
Four distinct layers: foundation model companies building the robot brains, data collection and annotation platforms providing the training fuel, open-source infrastructure lowering the barrier to entry, and managed data operations companies doing the actual field work. Each layer depends on the others. None of them are the same thing.
// Layer 01 — Robot Foundation Models
The companies training general-purpose robot brains. Their product is the model, not the data pipeline.
Skild's thesis is that there should be one brain for all robots - the Skild Brain, an omni-bodied foundation model that can control any hardware without prior knowledge of its form factor. Quadrupeds, humanoids, tabletop arms, mobile manipulators - same model, different embodiments.
Founded by Deepak Pathak and Abhinav Gupta, two of the researchers behind self-supervised learning, curiosity-driven exploration, and robot parkour. Their key insight: the data problem is too hard to solve purely with real-world teleoperation at the scale needed (trillions of examples, not millions). So they pretrain on internet video and large-scale simulation, then fine-tune with targeted real-world data for each customer deployment.
The data flywheel is core to the pitch: every deployment contributes new data back to improve the model across the fleet. Customers today include security and facility inspection, warehouses, manufacturing, data centers, and construction. Revenue grew from zero to ~$30M in a few months in 2025. Raised $1.4B at $14B valuation in January 2026, led by SoftBank with NVIDIA, Bezos Expeditions, and others.
Physical Intelligence - pi for short - is building general-purpose foundation models for dexterous robot manipulation. Their flagship models, π0 and π0.5, are vision-language-action (VLA) models that take camera input, proprioception, and natural language instructions and output continuous control signals for robot arms.
π0 was trained on data from seven different robotic platforms across 68 unique tasks. π0.5 pushed further toward open-world generalization - the ability to perform tasks in entirely new environments with new objects, not just settings seen during training. The core problem they're solving: not dexterity, but generalization. A robot that can fold shirts in one home but fails in another isn't useful at scale.
Both models are now available open-weights through Hugging Face's LeRobot. Scale AI lists Physical Intelligence as a customer for its Physical AI Data Engine.
// Layer 02 — Data Collection & Annotation Platforms
The infrastructure and tooling companies. They provide the platforms that power data collection, labeling, curation, and QA.
Scale AI is the largest and oldest data labeling company - founded in 2016 by Alexandr Wang, it built the foundational infrastructure for training data at scale, starting with autonomous vehicles and expanding to LLMs, robotics, defense, and enterprise AI. Revenue hit $2B in 2025. Meta acquired a 49% stake for $14.3B in June 2025.
Their physical AI push accelerated in late 2025 with the Physical AI Data Engine - a dedicated robotics data collection and annotation platform with a lab in San Francisco. Over 150,000 hours of robotics data delivered in 2025, with 10 new robotics customers onboarded including Physical Intelligence, Generalist AI, and Cobot. They're also integrating their platform into Universal Robots' UR AI Trainer.
Scale is the heavyweight - broadest operator network, deepest enterprise relationships, $2B+ in annual revenue. But the Meta deal created significant disruption: Google moved all key contracts off Scale, and competitors including Labelbox and Encord reported surges in demand from fleeing customers. The company's neutrality as a third-party data partner is now in question.
Encord is the fastest-growing annotation platform purpose-built for physical AI data. Their differentiator: a unified platform that handles the full lifecycle - data ingestion, curation, annotation, QA, and model alignment - across video, audio, LiDAR, RGB-D, radar, and sensor fusion, all in one environment. No tool fragmentation. No handoffs between systems.
In June 2025 they launched native 3D LiDAR and point cloud support, creating what they call the first unified Physical AI suite for robotics, autonomous vehicle, and drone teams. AI-assisted annotation can speed up labeling by up to 6x on temporal sequences. Customers include Archetype AI (2x faster annotation) and Pickle Robot (30% annotation accuracy improvement, 60% faster iteration cycles).
Raised $60M Series C in early 2026 led by Wellington Management with Y Combinator, CRV, and others, bringing total funding to $110M at a $550M valuation. Positioned as the technical alternative to Scale for teams that want a software-first, AI-native platform rather than a BPO-heavy operation.
// Layer 03 — Open-Source Infrastructure
Shared tools and datasets that lower the barrier to entry for the whole field.
Hugging Face's LeRobot is the open-source ecosystem for robotics AI - datasets, model weights, training scripts, and a standardized data format (LeRobotDataset: synchronized MP4/Parquet). It hosts thousands of community robotics datasets and is the distribution channel for major models including π0 and π0.5 from Physical Intelligence, and NVIDIA's GR00T N1.
LeRobot is solving the data fragmentation problem in robotics - the "ImageNet moment" that the field needs. Think of it as the Hugging Face Hub, but for robot demonstrations. Developers can fine-tune state-of-the-art robot policies with a few lines of code.
The Open X-Embodiment dataset (OXE) is the largest open cross-embodiment robot manipulation corpus - 22 robot platforms, 21 research institutions, 527 demonstrated skills across 160,000+ tasks. It's the common pretraining data baseline for most foundation models in the field.
Alongside DROID (564 real-world scenes across 52 buildings), OXE represents the floor of available public training data. Combined, they total roughly 5,000 hours of interaction data - far short of what production systems require.
// Layer 04 — Managed Data Operations
The companies that actually go out and collect, label, and deliver model-ready physical AI training data.
Field Motion is a physical AI training data infrastructure company - a managed service, not a platform. Trained operators, QA specialists, and real-world environment access. Core capabilities: teleoperation demonstrations, egocentric video with hand and finger tracking, dexterous manipulation sequences, force and contact data capture, sensor fusion datasets.
Data is designed for retargetability from the start - collected once, usable across multiple hardware platforms. Strong operator network in CIS and Slavic language locales. End-to-end pipeline management from first collection run to model-ready delivery. Clients include Apple, iSoftStone, and Defined.ai.
Alongside its platform tooling, Scale also operates a physical robotics data collection arm - a prototyping lab in San Francisco, plus a global contributor network. They handle custom collection from a wide range of embodiments in both lab and field settings, annotate with ML-assisted pipelines, and provide pre-built dataset libraries.
Scale's advantage here is volume and existing customer relationships. The tradeoff: it's a large-company operation, not a specialized field team. Bespoke collection needs - unusual environments, specific operator demographics, tactile/force data - are harder to execute at Scale's operating model.
How the layers connect
Foundation model companies like Skild AI and Physical Intelligence define what data they need. They either contract with data platforms (Scale AI's Physical AI Data Engine is a direct customer relationship with Physical Intelligence), use open-source datasets as pretraining baselines (OXE, DROID), or build proprietary collection operations in-house.
Annotation platforms like Encord sit in the middle - they don't usually collect data, but they provide the tooling for teams to label, curate, and QA whatever data gets collected. Some also offer managed annotation services (Encord has a labeling workforce) but their core product is software.
Open-source infrastructure (LeRobot, OXE) sets the baseline. Every serious player uses it for pretraining or model distribution - it's not a competitor to any of the above, it's shared infrastructure the whole field depends on.
The gap nobody has fully closed: managed field operations for specialized physical AI data collection. Scale AI has volume but operates like a large BPO. The robotics companies building foundation models need curated, diverse, in-the-wild data across specific hardware configurations, environments, and task types - with force and tactile data that lab setups rarely capture well. That's the operational problem that field-specialized teams exist to solve.
| Company | Layer | Core Product | Teleoperation | Annotation Tooling | Open Source | Force/Tactile Data |
|---|---|---|---|---|---|---|
| Skild AI | Foundation Model | Robot brain (Skild Brain) | — | — | — | — |
| Physical Intelligence | Foundation Model | VLA models (π0, π0.5) | — | — | ✓ | — |
| Scale AI | Data Platform + Ops | Data labeling & collection | ✓ | ✓ | — | Limited |
| Encord | Data Platform | Annotation + curation software | — | ✓ | — | — |
| Hugging Face LeRobot | Open Source | Dataset hub + model weights | — | — | ✓ | — |
| Open X-Embodiment | Open Dataset | Cross-embodiment training data | ✓ | — | ✓ | — |
| Field Motion | Managed Data Ops | End-to-end data collection | ✓ | ✓ | — | ✓ |
Need the data itself?
Field Motion handles collection, annotation, and QA for physical AI training data - teleoperation, egocentric video, manipulation sequences, force and contact capture. Managed end-to-end.