Low-power wireless networks have a hidden superpower, Time Slotted Channel Hopping protocol, or TSCH. This protocol creates a carefully orchestrated dance: devices take turns transmitting messages in precise time slots while hopping across frequencies to avoid interference. It's brilliant for reliability and energy efficiency, which is why it powers everything from smart factories to environmental sensors.

But there's a catch. These networks struggle when conditions change. A robot moves. Traffic patterns shift. A network node fails. Traditional algorithms can adapt time slot scheduling, but they're computationally hungry, far too expensive for devices running on milliwatts of power and a few kilobytes of memory.

This is the challenge that ADI researcher Martina Balbi tackled as part of her PhD thesis. Her question was deceptively simple: Can we bring AI to these tiny devices, to help us overcome these challenges, without breaking the protocol?

The obvious approach would be to train a neural network to replace the protocol scheduling logic entirely. Let the AI figure out scheduling, routing, everything. But Martina discovered something crucial: that approach fails.

In her experiments with scheduling, a basic neural network appropriate for the limited devices resources achieved 95% numerical accuracy at imitating an expert algorithm. Impressive, right? But only 61% of its outputs were actually valid according to protocol rules. In wireless communication, an invalid schedule doesn't just perform poorly, it creates collisions, wastes energy, and breaks the network.

This led to her central insight: Don't use AI to replace protocol logic. Use it as a constrained decision-support mechanism. Let learned models predict intermediate quantities, hints, recommendations, and scores while relying on deterministic mechanisms to enforce what's actually compliant with the protocol.

This hybrid approach is what we at ADI call Physical Intelligence: bringing AI from the abstract into the physical world, where millisecond decisions, imperfect sensing, and hard constraints define success or failure.

Martina demonstrated this philosophy across three distinct challenges and by contributing three distinct solutions in low-power wireless networks.

Making Scheduling Lightning-Fast

The first contribution tackled schedule computation. TSCH networks need to assign time slots and channels to each communication link, a complex optimization problem that traditional algorithms solve through elaborate computation.

Her approach: train a compact neural network to imitate an expert scheduler, then deploy it on an ADI MAX78000 microcontroller with a neural accelerator.

The results were stunning:

• Inference time fell to 104 microseconds from 201 milliseconds using the traditional algorithm.
• Energy consumption dropped to 5 microjoules from 1,606 microjoules.

That's 2,000 times faster and 300 times more energy efficient. A device that previously needed hundreds of milliseconds to compute a schedule can now do it microseconds. It’s like compressing a cross city drive into a street crossing and shrinking energy from a stadium floodlight to a laser pointer.

But here's the nuance: the model outputs are fed into a deterministic validator that ensures protocol compliance. The AI provides speed; the protocol provides correctness.

Adapting Routes Under Mobility

The second contribution addressed a harder problem: what happens when robots move? Static routing breaks down, packets get lost, and network performance degrades.

GENIAL is Martina’s centralized, genetic-algorithm–based cross-layer network management system. It jointly optimizes routing and TSCH scheduling to maintain high reliability and low latency in mobile networks composed of compute-constrained nodes. Under realistic conditions with 50 mobile nodes, GENIAL delivered 67% more packets than the standard approach while maintaining stable latency.

Then came GENIALNet: an imitation learning model that compresses GENIAL's behavior into a neural network, cutting orchestration time by 6 to 10 times. Interestingly, Martina’s research also revealed that when protocol constraints are tight enough, a well-designed heuristic can match the learned model's performance, a reminder that AI isn't always the answer, even when it works.

Flooding the Swarm

The third contribution, Plyo, tackles downstream communication in robot swarms. When a central controller needs to broadcast a command or update to hundreds of robots, the standard approach is simple: every robot rebroadcasts the message to its neighbors. This "flooding" is robust but wasteful. Robots with strong connections relay packets nobody needs, wasting energy, while the entire broadcast stalls until the hardest-to-reach robot finally receives it.

Plyo provides a mechanism that solves these challenges by selecting only the best-positioned robots as relays who will rebroadcast the message. It optimizes TSCH without breaking the protocol. Martina took things further by using AI for optimization, creating PlyoNet: a convolutional scoring network that learns to pick relays that maximize coverage to reach the farthest-off nodes. Because the protocol guarantees every choice is safe, the AI focuses entirely on performance — no deterministic validator needed.

Many of the challenges addressed by this work were surfaced by the OpenSwarm project. OpenSwarm, an EU-funded project, aims to enable novel, future energy-aware swarms of collaborative smart nodes with wide range benefits for the environment, industries, and society. To enable physical experimentation, OpenSwarm used DotBot, a low-power, 10-by-10-centimeter robot with the ultra-efficient ADI MAX78000 used as an AI accelerator. It is for such robots, and beyond it, a 1,000-robot swarm being set up at Analog Devices’ Limerick Catalyst facility. that GENIAL and Plyo were conceived. To learn more about ADI's contributions to the OpenSwarm, visit this article.

We believe that a test swarm like the one in Limerick, and those who will scale up beyond it in numbers and robot sophistication, will communicate more efficiently and robustly thanks to this work. First on earth, but also on the moon, where the physics are harder, with harsher, and with zero margin for error. The network architecture will still hold.