


Nonprofit AI institute that released MolmoAct, an open action-reasoning model for robot manipulation.



Built the RFM-1 robotics model for warehouse picking. In 2024 Amazon hired its founders and licensed its models.
Builds Dream-1, a robot foundation model trained on human video that turns a camera feed and a text instruction into actions for robots it has not seen, offered through an API.



Builds a hardware-agnostic autonomy stack and learned control models for humanoid and legged robots.
Trains GEN-series embodied foundation models on large-scale real-world robot manipulation data.



Research lab behind Gemini Robotics and Gemini Robotics-ER, vision-language-action models for robot control.



Provides tooling to teach robots new tasks through natural-language instructions and learned policies.



Jetson Thor robot compute, Isaac Sim/Lab, GR00T models and Isaac ROS.
Isaac 0.5, an open-weight robot model that reads video and outputs actions, and Mk1.5, a perception model for embodied agents sold by API.



Builds general-purpose robot foundation models (the π family of VLAs) for manipulation across robot embodiments; early versions were released as open weights (openpi).



Develops robot foundation models for dexterous, hand-level manipulation in industrial settings.
Develops Skild Brain, a general robot foundation model intended to run across different robot bodies.



Toyota-funded lab publishing Large Behavior Models and diffusion-policy research for dexterous manipulation.



NEO home humanoid and world models for robot learning.
Develops the DYNA-1 robot foundation model and runs it on manipulation tasks for commercial customers.



General-purpose humanoid robots (Figure 03) with the Helix vision-language-action model.



Deploys AI-driven robot systems for truck loading and parcel handling, built on its own manipulation models.



Develops field foundation models for robot autonomy and deploys them on robots in industrial and construction sites.



Bimanual robot data operations, the Loop demonstration-to-policy platform, and the CFG-1 manipulation model.



Builds robotics foundation models trained with physics simulation, maintains the open-source Genesis physics engine, and makes a sensor glove for capturing human demonstrations.



Open-source library of robot policies (including SmolVLA), datasets and tooling for training and sharing robot learning models.
The Robotics Map
A map of the robotics industry, from data and simulation to models, robots and deployments.
Filter Models
How to read this map ⌄
- Each station is one layer of the robotics stack, and its number counts the companies in it. The tiers read bottom-up, from data sources to what runs in the field.
- A line joins two layers that share companies, ones selling a real product in both. Its number says how many, and thicker lines mean more. Hover a line to see who; click it to keep the list open.
- +Zoom into a station, or click it, to open it. Companies only in that layer ring the station; companies that also work elsewhere sit on the line to that other layer.
- A company in four or more layers (today only NVIDIA) draws no lines. It sits inside every layer it works in instead.
Layers⌄
Busiest connections⌄
151 companies · 37 span more than one layer. Updated October 2026.
Built