Physical AI Company Directory 2026
Physical AI extends far beyond humanoid robots. The full stack includes simulation engines, sensor hardware, robot arms, world model researchers, foundation model companies, edge compute platforms, and deployment infrastructure. This directory maps 25+ companies across the entire physical AI ecosystem.
Use this as a reference when evaluating the landscape, looking for partners, or understanding where value is being created in the robotics and embodied AI supply chain.
Simulation and synthetic data companies
Simulation is the backbone of physical AI development. These companies provide the virtual environments where robots learn before touching the real world.
NVIDIA (Isaac Sim / Cosmos)
| Field | Details |
|---|---|
| Location | Santa Clara, California |
| Product | Isaac Sim, Cosmos, Omniverse |
| What it does | GPU-accelerated physics simulation for robot training, synthetic data generation, world model training |
| Pricing | Isaac Sim free for individual developers; enterprise licensing for Omniverse |
| Key customers | Most major robotics companies |
NVIDIA dominates physical AI simulation with Isaac Sim for robot-specific environments and Omniverse as the underlying 3D collaboration platform. Cosmos is their world foundation model trained on simulation and video data. The ecosystem is tightly integrated with their GPU hardware, creating a strong lock-in effect.
MuJoCo (Google DeepMind)
| Field | Details |
|---|---|
| Location | London, UK (DeepMind) |
| Product | MuJoCo physics engine |
| What it does | Fast, accurate contact-rich physics simulation for robotics research |
| Pricing | Free and open source (Apache 2.0) |
| Key users | Academic labs, DeepMind, research institutions globally |
MuJoCo (Multi-Joint dynamics with Contact) became free and open source after Google DeepMind acquired it in 2022. It remains the most popular physics engine for reinforcement learning research due to its speed and accuracy for articulated body simulation. Nearly every robot learning paper uses MuJoCo for benchmarks.
Genesis Embodied AI
| Field | Details |
|---|---|
| Location | Global (open-source project) |
| Product | Genesis simulation platform |
| What it does | Generative simulation platform combining physics engines with generative AI for automatic environment and task creation |
| Pricing | Open source |
| Key users | Research labs, startups exploring sim-to-real transfer |
Genesis is a newer entrant that combines physics simulation with generative AI to automatically create training environments. Rather than manually designing simulation scenarios, Genesis generates diverse environments procedurally, potentially accelerating the sim-to-real pipeline.
Gazebo (Open Robotics)
| Field | Details |
|---|---|
| Location | Mountain View, California |
| Product | Gazebo simulator |
| What it does | Robot simulation integrated with ROS 2 ecosystem |
| Pricing | Free and open source |
| Key users | ROS community, universities, startups |
Gazebo is the standard simulator for the ROS ecosystem. While less photorealistic than Isaac Sim, it integrates seamlessly with ROS 2 and is the default choice for teams already invested in the ROS stack. The recent Gazebo Harmonic release improved rendering and physics fidelity significantly.
Sensor and perception hardware companies
Robots need eyes, ears, and touch. These companies build the sensors that enable perception.
Intel RealSense / Orbbec
| Field | Details |
|---|---|
| Location | Intel: Santa Clara, CA; Orbbec: Shenzhen, China |
| Product | Depth cameras (D400 series, Femto series) |
| What it does | Structured light and stereo depth sensing for 3D perception |
| Price range | $200 to $800 per camera |
| Key users | Robot arm systems, AMRs, inspection systems |
RealSense (Intel) and Orbbec are the two dominant depth camera suppliers for robotics. After Intel announced winding down RealSense, Orbbec has gained significant market share with their Femto series offering comparable quality. These cameras provide the 3D point clouds that manipulation systems depend on.
Velodyne / Ouster (now Ouster)
| Field | Details |
|---|---|
| Location | San Francisco, California |
| Product | LiDAR sensors (OS series, digital LiDAR) |
| What it does | 360-degree 3D mapping for navigation and obstacle avoidance |
| Price range | $1,500 to $20,000+ depending on resolution |
| Key users | AMRs, autonomous vehicles, outdoor robots |
Ouster (which merged with Velodyne in 2023) makes digital LiDAR sensors used in mobile robots and autonomous vehicles. Their digital approach reduces cost compared to traditional spinning LiDAR while maintaining range and resolution. Essential for outdoor robots and large-scale warehouse navigation.
Prophesee
| Field | Details |
|---|---|
| Location | Paris, France |
| Product | Event-based vision sensors (neuromorphic cameras) |
| What it does | Captures only pixel-level changes at microsecond resolution, enabling ultra-low-latency perception |
| Price range | $500 to $3,000 (development kits) |
| Key users | High-speed manipulation, drone navigation, industrial inspection |
Prophesee builds event cameras that output data only when pixels change, rather than capturing full frames. This gives them microsecond temporal resolution with minimal data bandwidth. For robotics, this enables reactive grasping and high-speed manipulation that traditional cameras cannot support.
XELA Robotics
| Field | Details |
|---|---|
| Location | Tokyo, Japan |
| Product | Tactile sensing arrays (uSkin) |
| What it does | Distributed tactile sensors for robot fingers and grippers |
| Price range | Custom (development kits ~$2,000+) |
| Key users | Dexterous manipulation researchers, humanoid hand developers |
XELA makes tactile sensors that give robots a sense of touch. Their uSkin sensors measure 3-axis force distribution across surfaces, enabling robots to detect slip, measure grasp force, and identify object properties through touch. Critical for dexterous manipulation tasks.
Robotiq
| Field | Details |
|---|---|
| Location | Quebec City, Canada |
| Product | Force/torque sensors, adaptive grippers |
| What it does | End-of-arm sensing and grasping for collaborative robots |
| Price range | $3,000 to $15,000 (grippers); $2,000 to $5,000 (F/T sensors) |
| Key users | Universal Robots ecosystem, industrial cobot deployments |
Robotiq is the market leader in cobot accessories. Their 2F-85 and Hand-E grippers are installed on tens of thousands of collaborative robots worldwide. Their force/torque sensors enable compliant manipulation and contact detection essential for safe human-robot collaboration.
Robot arm companies
The robot arm market spans from $100 hobby arms to $100,000+ industrial systems.
Universal Robots
| Field | Details |
|---|---|
| Location | Odense, Denmark |
| Product | UR3e, UR5e, UR10e, UR16e, UR20, UR30 cobots |
| What it does | Collaborative robot arms for manufacturing automation |
| Price range | $25,000 to $60,000 |
| Key users | Manufacturing (SMEs), packaging, machine tending |
Universal Robots created the collaborative robot category and remains the market leader with over 75,000 cobots deployed globally. Their latest UR20 and UR30 models offer higher payload (20 kg and 30 kg) with longer reach. The UR ecosystem has hundreds of third-party peripherals and software integrations.
UFACTORY (xArm)
| Field | Details |
|---|---|
| Location | Shenzhen, China |
| Product | xArm 5, xArm 6, xArm 7, UFactory Lite 6 |
| What it does | Affordable collaborative robot arms for research and light industry |
| Price range | $5,000 to $16,000 |
| Key users | Research labs, startups, education, light manufacturing |
UFACTORY makes the xArm series, which offers 80% of Universal Robots functionality at 30% of the price. The xArm 7 (7-DOF) is popular in AI research labs for manipulation experiments. Their Python SDK and ROS 2 integration make them accessible for software-first robotics teams.
Hugging Face / Pollen Robotics (SO-100 / SO-101)
| Field | Details |
|---|---|
| Location | Paris, France |
| Product | SO-100, SO-101 (LeRobot ecosystem) |
| What it does | Ultra-low-cost open-source robot arms designed for AI research and teleoperation data collection |
| Price range | $100 to $300 (BOM cost for SO-101) |
| Key users | AI researchers, hobbyists, students |
The SO-100 and SO-101 are open-source robot arms designed specifically for collecting manipulation training data. Part of the LeRobot ecosystem (Hugging Face), they enable anyone to build a teleoperation setup for under $500. While not industrial-grade, they have become the default platform for academic manipulation learning research.
Franka Robotics (Franka Emika)
| Field | Details |
|---|---|
| Location | Munich, Germany |
| Product | Franka Research 3 (FR3) |
| What it does | Research-grade torque-controlled robot arm with sub-millimeter precision |
| Price range | ~$30,000 to $50,000 |
| Key users | Top robotics research labs (Stanford, MIT, CMU, etc.) |
Franka is the gold standard for manipulation research. Their torque-controlled 7-DOF arm enables compliant control and safe human interaction. Nearly every major robot learning paper from top labs uses Franka hardware. The FR3 added improved sensors and computing compared to the original Panda.
Flexiv
| Field | Details |
|---|---|
| Location | Santa Clara, CA / Shanghai, China |
| Product | Rizon 4, Rizon 4s, Rizon 10 |
| What it does | Adaptive robot arms with force control for contact-rich tasks |
| Price range | $30,000 to $70,000 |
| Key users | Automotive, electronics assembly, polishing/grinding |
Flexiv builds force-adaptive robot arms designed for tasks requiring constant contact with surfaces: polishing, deburring, assembly insertion. Their proprietary adaptive force control algorithms handle uncertainty in part positioning and surface geometry without extensive programming.
World model and foundation model companies
These companies build the AI brains that let robots understand and predict the physical world.
Physical Intelligence (Pi)
| Field | Details |
|---|---|
| Location | San Francisco, California |
| Product | Pi-0 foundation model for robotics |
| What it does | General-purpose robot foundation model that controls diverse robot hardware through a single neural network |
| Funding | $400 million+ (valued at $2.4 billion) |
| Key focus | One model for all robots, all tasks |
Physical Intelligence builds what they call a general-purpose robot foundation model. Their Pi-0 model can control different robot bodies (arms, mobile manipulators) across different tasks from a single pretrained model, fine-tuned for specific deployments. Co-founded by former Google Brain and Stanford researchers.
Skild AI
| Field | Details |
|---|---|
| Location | Pittsburgh, Pennsylvania |
| Product | Skild Brain (robot foundation model) |
| What it does | Scalable robot foundation model trained on massive simulation data |
| Funding | $300 million+ (valued at $1.5 billion) |
| Key focus | Scale-first approach to robot intelligence |
Skild AI, co-founded by CMU professors Deepak Pathak and Abhinav Gupta, takes a scale-first approach to robot foundation models. They believe that scaling data and compute for robot learning will yield the same breakthroughs that scaling produced in language models.
Covariant (now part of Amazon)
| Field | Details |
|---|---|
| Location | Emeryville, California (now Amazon Robotics) |
| Product | RFM-1 (Robotics Foundation Model) |
| What it does | Foundation model for warehouse manipulation, trained on billions of real-world picks |
| Status | Acquired by Amazon in 2024 |
| Key focus | Warehouse robot intelligence |
Covariant built RFM-1, a foundation model trained on data from billions of real warehouse picks across their customer deployments. After acquisition by Amazon, their technology feeds into Amazon Robotics operations. RFM-1 demonstrated the viability of foundation models trained on real robot data at scale.
NVIDIA Cosmos
| Field | Details |
|---|---|
| Location | Santa Clara, California |
| Product | Cosmos World Foundation Models |
| What it does | Generates synthetic video and physics-aware world simulations for robot training |
| Pricing | Available through NVIDIA ecosystem |
| Key focus | World models for physical AI |
Cosmos is NVIDIA’s world foundation model family, designed to generate physically plausible simulations of the real world. It produces synthetic training data for robots by predicting how the physical world evolves over time. Tightly integrated with Isaac Sim and the NVIDIA robotics stack.
Google DeepMind (RT-X, Gemini Robotics)
| Field | Details |
|---|---|
| Location | London, UK / Mountain View, CA |
| Product | RT-2, RT-X, Gemini Robotics |
| What it does | Vision-language-action models that combine language understanding with robot control |
| Status | Research and internal deployment |
| Key focus | Generalist robot policies via large multimodal models |
Google DeepMind’s robotics team builds vision-language-action (VLA) models that allow robots to follow natural language instructions by grounding them in visual perception and motor actions. Their RT-X initiative pools robot data across 22 institutions to train more general models. Gemini Robotics integrates their flagship LLM with embodied control.
Edge compute and deployment platforms
NVIDIA Jetson (Orin, Thor)
| Field | Details |
|---|---|
| Location | Santa Clara, California |
| Product | Jetson AGX Orin, Jetson Thor, Jetson Orin Nano 2 (announced) |
| What it does | Edge AI compute modules for robots |
| Price range | $200 (Orin Nano) to $1,999 (AGX Orin 64GB) |
| Key users | Virtually all robots requiring on-device AI inference |
The Jetson platform is widely used for robot edge computing. Jetson AGX Orin provides up to 275 TOPS of AI performance in a compact module, while Jetson Thor targets larger transformer-based robot policies. NVIDIA announced the 78-TOPS, 8 GB Jetson Orin Nano 2 on August 25, 2026; its module and developer kit are expected in H1 2027 and are not currently shipping. See the Nano 2 versus Orin Nano Super comparison for the practical deployment trade-offs.
Qualcomm RB series
| Field | Details |
|---|---|
| Location | San Diego, California |
| Product | RB3 Gen 2, RB5, RB6 |
| What it does | Robotics-specific compute platforms with integrated AI, connectivity, and sensor processing |
| Price range | $200 to $500 (development kits) |
| Key users | Consumer robots, drones, lightweight AMRs |
Qualcomm’s RB (Robotics Board) series targets smaller robots where power efficiency matters more than raw compute. Built on their Snapdragon platforms, these boards offer integrated 5G connectivity, camera ISPs, and AI accelerators in a compact form factor.
Infrastructure and tooling
Weights & Biases / Neptune AI
These MLOps platforms are widely used in robot learning for experiment tracking, model versioning, and training run management. Robot learning teams use them to track sim-to-real transfer experiments and compare policy performance across hardware variations.
ROS 2 (Open Robotics / Intrinsic)
The Robot Operating System remains the standard middleware for connecting sensors, actuators, planners, and AI models within a robot. Now maintained partly by Intrinsic (an Alphabet company), ROS 2 provides the communication backbone for most non-proprietary robot systems.
Foxglove
| Field | Details |
|---|---|
| Location | San Francisco, California |
| Product | Foxglove Studio (robot data visualization) |
| What it does | Web-based visualization and debugging for robot data streams (ROS, MCAP, custom) |
| Pricing | Free tier; paid plans from $50/month |
| Key users | Robotics engineering teams debugging sensor data and robot behavior |
Foxglove provides modern visualization tools for robot developers, replacing the aging RViz tool. Their web-based studio displays point clouds, camera feeds, joint states, and maps in a collaborative environment. Supports ROS 1, ROS 2, and custom data formats.
Company count by category
| Category | Count | Notable players |
|---|---|---|
| Simulation | 4+ | NVIDIA, MuJoCo, Genesis, Gazebo |
| Sensors | 5+ | Orbbec, Ouster, Prophesee, XELA, Robotiq |
| Robot arms | 5+ | Universal Robots, UFACTORY, Franka, Flexiv, SO-101 |
| World models / Foundation models | 5+ | Physical Intelligence, Skild, Covariant/Amazon, NVIDIA Cosmos, DeepMind |
| Edge compute | 2+ | NVIDIA Jetson, Qualcomm RB |
| Tooling / Infrastructure | 3+ | ROS 2, Foxglove, W&B |
How to use this directory
If you are building a robot system, map your needs to these layers:
- Hardware layer: Choose arms (UFACTORY for budget, Franka for research, UR for production) and sensors (Orbbec for depth, Ouster for LiDAR).
- Simulation layer: Start with MuJoCo for learning research or Isaac Sim for photorealistic training.
- AI layer: Evaluate Physical Intelligence or Skild for foundation model approaches, or train your own policies using open-source frameworks.
- Compute layer: Default to Jetson Orin for most applications.
- Tooling layer: Use ROS 2 for middleware and Foxglove for debugging.
Frequently asked questions
What is the physical AI stack? The physical AI stack includes all layers needed to build intelligent robots: hardware (arms, sensors), simulation (virtual training environments), AI (foundation models, policies), compute (edge processors), and deployment tooling (ROS, visualization).
Which physical AI companies are publicly traded? NVIDIA (NVDA), Qualcomm (QCOM), Intel (INTC), and UBTech (9880.HK) are publicly traded. Most others are venture-backed or corporate subsidiaries.
How much funding has gone into physical AI? Physical AI companies raised over $18 billion in the first half of 2026 alone. The sector has attracted more than $50 billion in cumulative venture investment since 2020.
What is the most important layer of the physical AI stack? There is no single most important layer, but simulation and foundation models are where the most investment is concentrated in 2026. Hardware remains essential but is becoming more commoditized at lower price points.
Can I build a physical AI system with only open-source tools? Yes. MuJoCo (simulation), ROS 2 (middleware), SO-101 (hardware), and open-source VLA models provide a complete stack at minimal cost. Performance will trail proprietary solutions but is sufficient for research and prototyping.
Which companies are likely acquisition targets? Sensor companies (Prophesee, XELA), smaller arm manufacturers (Flexiv), and foundation model startups (Physical Intelligence, Skild) are frequently discussed as acquisition targets for larger tech companies expanding into physical AI.
Sources
- NVIDIA Isaac Developer Platform — simulation, training, and deployment tools
- MuJoCo GitHub (Google DeepMind) — open-source physics engine (Apache 2.0)
- Physical Intelligence raises $400M – official blog — Pi-0 model announcement and funding
- Universal Robots – UR20 product page — 20 kg payload cobot specifications
- UFACTORY xArm Developer GitHub — open-source SDK for xArm robots
- Foxglove Studio – official site — robot data visualization and debugging
- Raspberry Pi AI HAT+ product page — Hailo-8 accelerator for robotics perception
- NVIDIA Jetson Orin Nano Super – Buy page — edge AI compute modules and pricing
- NVIDIA Jetson Orin Nano 2 announcement — announced status, specifications, and expected H1 2027 availability
- Franka Robotics – official site — FR3 torque-controlled research arm
- Flexiv Robotics – official site — adaptive force-controlled robot arms