Toward a humanoid foundation model
with a world model inside.
Kepler is our foundational model for physical intelligence, in development across every body we ship. It plans slowly, reacts fast, and learns from every hour the fleet operates.
01
Architecture
Think slow. React fast.
Two systems, one brain. A deliberative cerebrum plans at a few hertz; a reflexive cerebellum acts at hundreds. Plans flow down, state flows back up — at every tick.
High-level planner. Vision-language reasoning, task decomposition, long-horizon intent.
Reflex layer. Balance, contact response, force control — faster than thought.
→ actuators
02
World Model
Designed to predict contact before motion.
Kepler v1.0 targets a 10B-parameter world model trained on vision, force, tactile, teleoperation, simulation, and real-robot rollout data.
- Inputs
- Vision · Force-torque · Tactile
- Predicts
- Action outcomes · Contact dynamics
- Scale
- 10B parameters
- Training
- Pretraining + RL post-training on real-robot rollouts
03
Data Flywheel
Data pours in. Intelligence ships out.
Real production floors in Japan, our own hardware collecting sensor and video data, and a model stack that never stops training. Three moats, one loop.
26,000,000
hours of targeted real data
expected hours of training data — 2027 alone
04
Compute
Meet HISUI.
Our dedicated training cluster. Kepler pretrains and post-trains on it around the clock — and it grows with the fleet.
GPU
B300 × 640 GPUs
NVIDIA B300 — 80 nodes (expected)
Storage
24 PB
Fleet sensor & video data (expected)
05
Join