Grounded Superintelligence
Grounded Superintelligence (GSI) is building a human-centric robotics platform that trains robot policies from in-the-wild human behavior data rather than teleoperation. The company's RoboCap suite—a 250g six-camera hat with synchronized wrist cameras—captures egocentric manipulation data, which the Grounded API converts into training tokens using state-of-the-art SLAM, depth estimation, and hand tracking models.
- Robotics
- Artificial Intelligence
- Industrial Automation
- Manufacturing / Industry 4.0
- Hardware
Funding
Founders
Founder details are not available yet.
Product
Problem
Current approaches to robot learning rely heavily on teleoperation, which requires trained human operators, working robots, and is constrained by deployable workspaces and tasks. This method is too slow and expensive to reach the data scale needed for general-purpose physical intelligence, and robot data collected on one morphology is incompatible with others, failing to reflect the true distribution of human labor.
Solution
Grounded Superintelligence is building human-centric robotics technology that learns from human behavior data collected in the wild, rather than from robot teleoperation. The company has demonstrated how to transfer data from smart glasses worn by humans and tactile human data into robot policies, enabling robots to learn with a sense of touch. This approach promises orders-of-magnitude gains in sample efficiency and data collection speed compared to current methods from Physical Intelligence, Figure, Tesla, and Google DeepMind. The platform is vertically integrated over operations, hardware, and AI models, allowing the company to develop a human foundation model that scales with human data and transfers to humanoid robots.
Target Audience
Primary customers are industrial automation buyers, including factories and household robotics providers, as well as data vendors, business owners with workforces performing manual tasks, and researchers seeking high-fidelity hand manipulation data for human-to-robot learning experiments.
Features
- Transfer of in-the-wild human data from smart glasses into robot policies
- Integration of tactile human data into robot policies, giving robots a sense of touch
- RoboCap hardware: a 250g six-camera, dual-IMU hat with hardware-synchronized global-shutter cameras (1920×1080 at 30 Hz) and two wrist cameras, all wirelessly synchronized
- Grounded API: a self-serve application providing state-of-the-art 3D computer vision models for SLAM, depth estimation, and hand tracking, with GPU-accelerated SLAM beating ORB-SLAM3 and NVIDIA's cuVSLAM on public benchmarks
- Multi-view hand tracking pretrained on a proprietary corpus that tops all of Meta's released datasets, achieving sub-1cm accuracy
- Online self-calibration that reduces median front-pair epipolar residuals from 3.8 px to 0.56 px across 150 deployed devices
- Human foundation model architecture compatible with human data streams, distinct from robot-specific vision-language-action models
- Vertically integrated stack spanning operations, hardware, and AI model development