Nexa · Android
Not a human-shaped machine — a body matched to the human mechanism: tendon-driven muscle, full-body sensing skin, and a skeleton held to the same proportions and physical limits as ours.
Dexterous face
Expression is not a texture problem. Each muscle is studied for its motion effect, then a lattice structure is designed backwards from that effect to form a micro actuation unit — so expressions emerge from muscle, the way they do in a face.
Dexterous face system
Natural expression begins beneath the surface. Nexa maps facial muscles to lattice actuation units so fine movement emerges from anatomy rather than a surface animation.
Facial degrees of freedom
FACS action units
Perception system
A bionic sensing network spanning vision, touch, force, position, motion and physiological signal — fused, aligned, and resolved into a single body-state estimate.
System capabilities
Vision · touch · force · position · motion · physiology
Spatiotemporal synchronisation, cross-modal mapping
Pose · contact · load · balance · motion trend
Motion reconstruction, closed-loop feedback
Modalities and sensors
Touch / force
High-density tactile array · six-axis force sensor
Vision
RGB-D / stereo · first-person view
Audition
Microphone array · bone conduction
Position / deformation / motion
IMU / positional markers · deformation sensors
Physiological
Heart rate · respiration · EMG · surface temperature · thermal distribution
Unified body-state output
Face
Micro-expression / EDA / thermal
Perception fuses into one unified body state — knowing where it is, how it is touched, and how to respond.
Biofidelity Benchmark
Nexara and the Society for Android Science jointly define a six-layer Biofidelity Benchmark — turning “human-like” from a marketing adjective into a standard you can measure layer by layer.
What cannot be measured cannot be built.
Each layer names what it evaluates, the question it answers, and the metrics it answers it with. Layers are measured independently — a system can hold at L3 and fail at L5.
| Layer | Evaluates | Core question | Example metrics |
|---|---|---|---|
| L1 Morphological Fidelity | Body form, skeleton, anatomical structure | Does it look like a human? | Body proportion · Skeletal structure · Joint axes · Muscle layout · Facial proportion |
| L2 Biomechanical Fidelity | The motion system | Does it move like a human? | Range of motion · Gait · Postural stability · Joint torque · Energy consumption · Motion smoothness |
| L3 Biophysical Fidelity | Skin, soft tissue, temperature, contact | Does it feel like a human? | Young's modulus · Rebound · Surface temperature · Coefficient of friction · Skin texture |
| L4 Sensorimotor Fidelity | Perception and bodily feedback | Does it sense its own body the way a human does? | Tactile density · Visual coverage · Proprioception · State estimation · Response latency |
| L5 Expressive Fidelity | Expression, gaze, voice, posture | Can it express human emotion? | Expression accuracy · Vocal affect · Gaze naturalness · Lip-sync · Gestural semantic consistency |
| L6 Social Fidelity | Long-horizon interaction | Is it treated as a real person? | Trust score · Uncanny valley · Likability · Persona consistency · Long-term acceptance |
L1 to L6 runs from looking like a human to being treated as one. Fidelity accumulates; a layer is only meaningful once the layers beneath it hold.
Why it exists
Conventional robotics scores itself on task performance — faster, stronger, more efficient. An android is scored on human fidelity. Those are different objective functions, and they need different yardsticks.
Protein structure had the PDB. Vision had ImageNet. Without a shared, published measure, every claim about human likeness is unfalsifiable — and a field of unfalsifiable claims never becomes a science.
The same six layers that let researchers compare systems let a licensee write verification criteria into a contract. One standard, two uses.
Method
The Human Foundation Dataset feeds two consumers: benchmark evaluation and model training. The two are isolated by purpose, so a system is never scored against the data it learned from.
Nexara does not sell to consumers. What we build is licensed to a small number of partners.