/ NEXT SCALE FRONTIERMI-NXT-10M
10M
frontier
SPEC / NEXT-GEN DATA SCALING
AXIS X / EXPERIENCE SCALE
AXIS Y / EXPERIENCE DENSITY
AXIS X / EXPERIENCE SCALE
AXIS Y / EXPERIENCE DENSITY
01
The Next Frontier / 10M
Scale the hours. Increase what each hour can teach.
The path to 10M is two-dimensional: a greater volume of human experience, and denser physical information within each experience. Pose, geometry, object states, contact, force, and temporal structure turn raw video into learnable physical intelligence.
Environment
Task
Subtask
Instruction
/ ACCUMULATED EXPERIENCEMI-SCL-2.0M
2.0M
hours
SPEC / EXPERIENCE SCALE
ACCUMULATED / 2,000,000 HOURS
STATE / FOUNDATION
ACCUMULATED / 2,000,000 HOURS
STATE / FOUNDATION
02
Experience Scale / 2.0M Hours
Two million hours
is a foundation,
not a finish line.
is a foundation,
not a finish line.
We have accumulated two million hours of real-world experience. At this scale, the advantage is not volume alone, it is the ability to curate, search, balance, and assemble the right mix of experiences for each model.
Mi / Sep / 2026
/ MONTHLY CAPACITYMI-CAP-450K
450K
hours / mo
SPEC / PRODUCTION THROUGHPUT
CAPACITY / 450,000 HOURS / MONTH
NETWORK / DISTRIBUTED
CAPACITY / 450,000 HOURS / MONTH
NETWORK / DISTRIBUTED
03
Production Capacity / 450K
Research-grade data
at industrial scale.
at industrial scale.
Our distributed collection and processing system is built to produce up to 450,000 hours of video per month, while keeping task design, capture, QA, and structured outputs connected in one production loop.
/ SCENE DIVERSITYMI-DIV-50000
50000
scenes
SPEC / ENVIRONMENT DIVERSITY
COVERAGE / 50000+ SCENE TYPES
MODE / NATURAL VARIATION
COVERAGE / 50000+ SCENE TYPES
MODE / NATURAL VARIATION
04
Scene Coverage / 50000
Not just a filter.
Diversity is built in.
Diversity is built in.
Our data spans 50,000 different scenes across homes, kitchens, workshops, labs, retail, warehouses, outdoors, and more. With the clutter, lighting, viewpoints, people, and edge cases that real deployment brings.
/ EXPERIENCE MEASURESMI-SIG-009
9
measures
SPEC / MULTIMODAL OUTPUTS
INPUT / EGOCENTRIC VIDEO + SENSORS
OUTPUT / ROBOT-READY SIGNALS
INPUT / EGOCENTRIC VIDEO + SENSORS
OUTPUT / ROBOT-READY SIGNALS
05
Experience Density / 9 Measures
Beyond raw video. Enriched with
physical modalities.
physical modalities.
Each hour can become a richer training signal. Our perception and sensing stack turns raw experience into structured information about actions, geometry, motion, contact, force, and space.
Signal / perception modules →09 active outputs
01
Video Understanding
Actions, activities, temporal understanding
02
Depth
3D structure and geometry
03
Object Tracking
Instances and trajectories
04
Hand Poses
Hand state and manipulation
05
Upper Body Poses
Body kinematics and motion
06
Tactile Force
Contact, pressure, and force signals
07
Localization
Camera and robot pose in the world
08
Segmentation
Scene and object labels
09
Point Cloud
Dense 3D representations
/ ACTION COMPLEXITYMI-ACT-003
3
levels
SPEC / HUMAN EXPERIENCE DATA
RANGE / EASY → HARD
MODE / REAL-WORLD
RANGE / EASY → HARD
MODE / REAL-WORLD
06
Action Coverage / 3 Levels
From picking to packing,
we covered every manipulation.
we covered every manipulation.
We capture the full spectrum of actions, from easy to hard.
Our data goes beyond clean demonstrations, exposing models to real-world ambiguity, recovery scenarios, complex coordination, and long-horizon tasks.
Our data goes beyond clean demonstrations, exposing models to real-world ambiguity, recovery scenarios, complex coordination, and long-horizon tasks.
Complexity ProfileEasy / Medium / Hard
Easy
Single-step
Repeatable
Repeatable
Medium
Multi-step
Coordinated
Coordinated
Hard
Long-horizon
Contact + Recovery
Contact + Recovery
RESEARCH / LATEST TECH NOTEMI-RSCH-001
Research / Latest
Beyond 1M Hours: The Data Scaling Laws of Physical AI.
Scaling physical intelligence is not only about collecting more hours. The next frontier is to scale experience volume and increase the learnable physical structure contained inside every hour.
READ TECH NOTE ↗LATEST / SEP 2026PHYSICAL AI / DATA SCALING