Highlanders
Kepler builds foundational AI models for physical intelligence, using its own robot hardware to capture real-world sensor data that continuously improves its world model. The company targets industrial applications across automotive, logistics, and manufacturing, with over 5,000 pre-orders and a planned 10B-parameter model trained on vision, force, and tactile data.
- Robotics
- Artificial Intelligence
- Industrial Automation
- Manufacturing / Industry 4.0
- Hardware
Funding
Founders
Founder details are not available yet.
Product
Problem
Industrial robots and autonomous systems often struggle to operate reliably in unstructured, real-world environments because their AI models lack deep understanding of physical dynamics, contact forces, and long-horizon task planning. This limits their deployment in dynamic settings like factories, warehouses, and logistics hubs where adaptability and safety are critical.
Solution
Kepler is developing a foundational world model for physical AI, designed to give robots a unified understanding of their environment. The company builds its own robot bodies to create a closed data flywheel: every deployed unit continuously collects sensor, video, and force-torque data from real production floors, which is fed into a training pipeline. This data trains a two-system architecture—a deliberative "cerebrum" for high-level planning at a few hertz and a reflexive "cerebellum" for fast, contact-level control at hundreds of hertz. The result is a model that plans slowly, reacts fast, and improves with every hour of fleet operation, enabling more robust and adaptable physical intelligence.
Target Audience
Primary customers are industrial enterprises in automotive, logistics, and manufacturing seeking to deploy humanoid or general-purpose robots for complex physical tasks in real production environments.
Features
- Two-system neural architecture: a deliberative cerebrum (vision-language reasoning, task decomposition) and a reflexive cerebellum (balance, contact response, force control) operating at different frequencies
- Kepler v1.0 targets a 10B-parameter world model trained on vision, force-torque, tactile, teleoperation, simulation, and real-robot rollout data
- Dedicated training cluster "HISUI" with 640 NVIDIA B300 GPUs and 24 PB of fleet sensor and video data storage
- Proprietary hardware designed in-house to own the feedback loop, with one software stack supporting multiple robot bodies
- Data flywheel that converts every operator interaction and field hour into training data, with 26 million expected hours of targeted real data in 2027 alone