Innatera's Pulsar is 2.8 by 2.6 millimetres. It runs a spiking neural network of roughly 200 neurons inside a robotic fingertip, and it classified paper, plastic, foam and wood with 98% accuracy as the finger moved across them.

That number matters because of what is missing. No GPU. No cloud round-trip. No central model deciding what the hand already knows.

The deployment, disclosed by Innatera on 19 October, is one of the first production examples of neuromorphic silicon doing real work in a robot rather than in a demo video. Analog Devices built the finger. Pulsar runs the classifier. The reported power gain over a conventional neural processing unit is 25 times.

💡
Why a smaller brain wins at the edge
Reflex loops stay in the hand, so slip correction does not wait on a central processor.
Inference fits inside a milliwatt budget, which matters for untethered machines.

A reflex that never leaves the hand

Most robot stacks centralise perception. Cameras, grippers and force sensors feed a large model on a central GPU, which sends instructions back to the motors. It works, and it is expensive. A humanoid running a datacentre-class accelerator draws tens to hundreds of watts before it lifts anything.

The alternative is to put a small classifier next to the sensor. Pulsar combines three fabrics on one die: a neuromorphic core of silicon neurons and synapses, a conventional deep-network accelerator, and a RISC-V subsystem for housekeeping. It reads inertial units, microphones, radar and low-resolution cameras directly.

That is the difference between thinking about a grip and adjusting it. When an object starts to slip, the correction has to happen now, not after a scheduler grants time on the central model.

Eighty million parameters became a few thousand

Innatera's own account puts the compression in perspective. A material-classification workload that previously ran an 80-million-parameter model on GPUs reached equivalent accuracy on Pulsar with a model in the kiloparameter range, at sub-milliwatt power, according to Innatera's account of a Sensors Converge panel in May. The company says it got there by matching the architecture to the signal rather than shrinking a large model after the fact.

The industrial proof is less glamorous and more convincing. Consultancy 42 Technology built a condition-monitoring module around Pulsar that watches motor vibration and flags emerging faults within seconds. It runs on 600 microwatts or less, with an expected system cost below $10, and it has been shown at CES, Embedded World and Edge AI London.

Vendor figures deserve scepticism. Innatera claims up to 100-fold faster processing and 500-fold lower power than conventional AI processors. Treat those as ceiling numbers from a friendly benchmark, not field averages.

For defence robotics the argument is less about efficiency than about what happens when the link disappears. A ground vehicle or drone that loses its radio, or operates under jamming, cannot wait for a remote model to answer. Local inference keeps the machine responsive when the network is the first thing to fail.

That logic is pulling defence primes toward the same silicon. Safran Electronics & Defense is exploring neuromorphic cameras with startup Neurobus for drone detection, a sensing problem where event-driven pixels can be far cheaper than frame-based video. BrainChip shipped an SPI co-processor on 22 October that lets existing microcontrollers add on-chip learning without a redesign. The pieces are converging on the same placement: intelligence at the sensor, not at the centre.

The caveat is honest. Pulsar's public deployments are industrial and consumer. Neither Innatera nor its partners will name a fielded defence product built on it. The strategic case remains a projection from proven sensing work, not a procurement record.

What to watch

Three signals will show whether this is a component transition or a curiosity. First, whether more robot makers publish Pulsar-class deployments with accuracy and power numbers attached. Second, whether defence programmes move from camera trials to named sensor contracts. Third, whether the toolchain gets boring. Innatera's developer portal promises the Talamo SDK worldwide, and boring tools are how a chip stops being a science project.

For now the interesting shift is spatial. The model is getting smaller, cheaper and closer to the metal. As we wrote in our piece on Overland AI's ground-autonomy contract, the hard part of autonomy is rarely the model itself; it is everything around it that has to work when the environment refuses to cooperate.

Innatera: robotics need a softer kind of compute
Innatera's own account of the Analog Devices tactile-finger deployment, with the 98% classification result, the roughly 200-neuron network and the 25x power comparison.
Innatera
The primary disclosure for the deployment claim, written by the vendor — read it with that in mind.
42 Technology: 600 microwatts for always-on monitoring
A working Pulsar-based demonstrator that catches motor faults in seconds at a system cost below ten dollars — the most concrete cost and power figures in this space.
42 Technology
Independent engineering consultancy, vendor silicon, numbers you can sanity-check.
Pulsar inside the power-constrained edge
Trade-press walkthrough of the Pulsar architecture and the three-fabric design, useful background on why a microcontroller can act as a sensor-local classifier.
Embedded Computing Design
Best available structural description of the chip for readers who want the block diagram in prose.