LFX-02

One model.
Any machine.

01Architecture

One model.
Four faculties.

LFX-02 unifies body, environment, coupling, and intent in a single foundation. Weights are shared. Competence in one faculty propagates to the rest.

01BodySelf-Model · φ
02EnvironmentWorld-Model · Ω
03CouplingInteraction Policy · π
04IntentMission Compiler · τ
02Embodiments

Any platform.
Zero retraining.

Four embodiment classes in service. The architecture is platform-agnostic. The catalog expands with every deployment.

Humanoid · CL-01 Humanoid blueprint

Humanoid

CL-01

Bipedal locomotion. Dual-arm manipulation. Contact-rich operation at human scale.

Drone · CL-02 Drone blueprint

Drones

CL-02

Multirotor, fixed-wing, hybrid. Autonomous flight envelopes. Swarm coordination at scale.

Rover · CL-03 Rover blueprint

Rovers

CL-03

Wheeled and tracked ground systems. Off-road traversal. Persistent autonomy in the field.

Manipulator · CL-04 Manipulator blueprint

Arms

CL-04

Industrial and dexterous manipulators. Tool use. Precision assembly. Bimanual coordination.

03Method

How it learns.
Four stages. Fixed order.

Competence is acquired in sequence. Each stage compounds the last. The result: one model that inhabits any body, in any environment, with intent.

Stage 01

Body

Self-Model · φ

The model first maps its own embodiment. Joint limits. Actuator dynamics. Sensor topology. Kinematic chains. An internal forward model of the machine it inhabits.

InputProprioception, IMU, Encoders
Wall-time~ 40 minutes
Stage 02

Environment

World-Model · Ω

Then it builds the world around it. A persistent spatial and semantic map: geometry, objects, affordances, agents. Composed from exteroceptive sensors. Updated in real time.

InputRGB, Depth, LiDAR, Radar
Wall-timeContinuous
Stage 03

Body ↔ Environment

Interaction Policy · π

The model then couples the two. It learns the dynamics of contact: how its body deforms the world, how the world resists, how force propagates. The sensorimotor loop closes.

Inputφ ⊗ Ω
Wall-time~ 4 hours
Stage 04

Goal-Driven Task

Mission Compiler · τ

Finally, intent. Natural-language and structured objectives compile into long-horizon plans. Executed by π. Continuously re-grounded against the live world model.

InputMission specification
Wall-time42 ms / decision

We do not build robots. We build the intelligence every robot runs on.

Member of
NVIDIA Inception NVIDIA Inception
Google Research Google Research
Team from
Palantir Palantir
Bethesda Studios Bethesda Studios
AMD AMD

Give it a body.
It does the rest.

LFX-02 is available to qualified research and commercial partners. Request access to integrate the model with your platform.