NexSim · Simulation

Where Nexa learns before it acts.

Physiologically accurate simulation of a real biofidelic body, fast enough to train at scale.

Why now

Simulation crossed a computational threshold

Physiologically accurate musculoskeletal simulation — built on real biomechanical models rather than rigid-body approximations — used to be far too expensive for reinforcement learning at scale. It is not any more. A high-fidelity body can be modelled and trained in simulation before any hardware exists.

Why it has to be biofidelic

A rigid-body simulator trains the wrong intuitions

Train on rigid-body data at scale and you get something impressive: real, hard-won intuitions about force and balance. Just not ours — intuitions about being that machine. More data does not fix it, because the distortion is systematic rather than noise. It only makes the system more confident in being wrong.

Simulating a body that pulls rather than pushes, that carries redundant opposing muscle groups and distributed sensing, is the only way the training transfers to one.

The loop

Where NexSim sits

Four products, one loop. NexSim is the step between capturing what a person did and running a model inside a body.

  1. 01

    Build Nexa

    The most biofidelic android — the only kind whose competence can be measured against a person's directly, because it is the only one under the same constraints.

  2. 02

    Capture the Human Foundation Dataset with NexCap

    Not what someone did, but how a body learned to do it.

  3. 03

    Train Nexus in NexSim

    The Human Foundation Model, trained against a physiologically accurate body rather than a rigid approximation of one.

  4. 04

    Run Nexus inside Nexa

    And measure everything against the Biofidelity Benchmark.

The argument in full

Why a body that matches the human mechanism — rather than merely resembling it — is the precondition for general intelligence, not an application of it.