Intelligence that grows

Like a kid learning to catch a ball, the brain rewires itself with every try. No instruction manual. No giant training run. A million neurons, learning the way yours did.

Today's robot brains burn power like a space heater. OSCEN is built to sip it like a phone charger. Published spiking-chip benchmarks show up to 109x less energy per inference (Blouw et al. 2019), microjoules instead of millijoules. OSCEN is built to run on that class of chip, so robots can work all day on one battery.

Like a baby becoming a toddler becoming a kid, OSCEN moves through five stages of growth. Each one rewires what it can learn next. Old skills stay even as new ones come in.

Cameras, microphones, motors, speakers. Built body-agnostic: a new body is a manifest and a plugin, with no brain changes. The brain is the product. The body plugs in.

VLAs are trained on millions of demonstrations, then deployed static. They can only do what they've already seen. A new door handle, an unexpected obstacle, a different lighting condition. Anything outside training data is a failure. Retraining costs hundreds of thousands of dollars and weeks of GPU time.

Every robot action requires a full forward pass through billions of parameters. At 10Hz control rate, that's 30 billion matrix multiplications per second, continuously. A warehouse fleet of 100 robots consumes 5-20kW just thinking. OSCEN targets under 5W total through sparse, event-driven activation. The same power as an LED bulb.

Most VLAs are too large to run on-device efficiently. Physical Intelligence requires cloud. When your robot arm is holding a baby or a surgical tool, 200ms of network latency isn't an engineering tradeoff. It's a liability. OSCEN is designed to run entirely on-edge. No internet, no latency, no data leaving the device.

Fine-tune a VLA on new tasks and it forgets old ones. This is a fundamental limitation of gradient-based learning in large models. OSCEN's homeostatic synaptic scaling, myelination, and identity tagging (Patent Claim 6) are designed to prevent this, preserving learned skills as new ones are acquired. A retention experiment is running now.