IntuitoContact
01Intuito
Physical reasoning for robots.
Give a robot a goal, not a demonstration. Our chip and our intuition model work out every move with physics.
02The problem
Today's robots memorize.
Each task costs hours of human demonstrations and costly training. Then, in the real world, whatever they were not shown makes them fail without warning.
03The reason
Thinking was too expensive.
Simulating what will happen never fit inside a robot, so robots memorize instead. Our chip, the WPU, is designed to make that thinking almost free.
04WPU · world lanes
It imagines many futures before every move.
Each lane simulates one possible world, with its own physics. The robot picks the move that works in all of them.
05WPU · contact solver
Exact where it matters: touch.
Grasping, pushing, holding. The WPU is designed around contact, the part of physics where GPUs fall behind.
06WPU · intuition network
Intuition that grows with experience.
What the robot solves by thinking becomes instinct, designed to improve with practice and be shared across robots. The network proposes; physics checks every move.
07WPU · practice
When it rests, it practices.
Off duty, the same chip is designed to replay what was hard across many simulated worlds. That practice becomes sharper intuition, for every robot.
08WPU · the whole die
Same answer, every time.
Designed to give identical results on the robot and in the data center, so any failure can be replayed and fixed.
09Measured, not assumed
It measures the world. It does not assume it.
It is designed to touch, nudge and weigh: mass, friction and gravity become numbers it measures. When reality surprises it, it slows down and corrects.
10The product
The whole brain, not one piece.
Language models understand the order. Vision finds the object. Our chip decides each movement, alone inside the control loop.
- Taskslanguage model
- Seeingvision
- Actionour chip and method
- Bodymodel of the robot
- Practiceimproves at rest
- Supervisionsays when no plan is safe
11The difference
Memorizing vs. reasoning.
12Any robot
One goal, any body.
A task is stored as a goal, not a movement. A new robot should need a model of its body, not new training.
Targets
What this changes.
These are our targets, not results. Today the method runs one task in simulation; the targets will be measured against teleoperation and against training in simulation.
- 01
No demonstrations
No teleoperation and no training per task. A task is given as a goal.
- 02
Minutes, not weeks
Our target for a new task, from describing it to the first success.
- 03
Low-cost robots
Precision is meant to come from computation, not from expensive hardware.
- 04
One brain, many jobs
A new task should need only a new goal. Long tasks become chains of goals.