Meet Sniffbot, a robot that smells with a real locust antenna

01The question

Finding where a smell is coming from matters for spotting gas leaks, explosives, drugs or spoiled food. Today's options all have drawbacks. Laboratory instruments are extremely precise but slow and not portable. Electronic noses are portable but less sensitive. Sniffer dogs are excellent, but they take a long time and a lot of money to train, and each dog learns only a few target smells.

Insects, on the other hand, have remarkably sensitive antennae that respond to a huge variety of odors. The researchers wanted to know whether a real insect antenna could serve as a robot's nose, and whether that robot could find a smell's source indoors, where there is no wind to carry the scent in a trail.

02What they did

A team at Tel Aviv University built Sniffbot on a small robot car about the size of a shoebox (21 × 15 × 11 cm) and gave it three layers. The sensing layer holds an antenna taken from a desert locust. When odor molecules land on it, the antenna produces tiny electrical signals (an electroantennogram) that a miniature amplifier picks up.

The sniffing layer uses a small pump and valve to pull in short puffs of air, like an animal sniffing. This concentrates the odor and stops the antenna from getting "used to" a smell (habituation). The deciding layer is a Raspberry Pi mini-computer that reads the signals and either chooses where to drive or uses a machine-learning model to name the smell.

The team tested Sniffbot in a closed 3.8 × 4.8 m room with the air conditioning off, so the air was almost perfectly still. Using lemon oil as the target, they compared three search strategies, with 15 runs each and a limit of 50 moves per run.

03What they found

Sniffing made a real difference. With active sniffing, the signal got steadily weaker as the robot moved farther from the odor, which is exactly the clue a robot needs to navigate. Without sniffing, the signal said almost nothing about distance.

A new search strategy the team designed, called Trident, reached the odor source in 93% of runs, compared with 53% for a spiral search and 30% for a random-wandering strategy inspired by how E. coli bacteria swim.

When asked to tell smells apart, Sniffbot correctly picked out benzaldehyde (which smells like almonds) and an odorless blank in over 90% of trials, and citronellol (a rose-like scent) in 75%. Random guessing would score 33%.

04Why it matters

Following a smell without wind is one of the hardest problems in odor tracking. It's also the situation inside warehouses, homes or collapsed buildings. Sniffbot shows that a biohybrid machine, part living tissue and part electronics, can handle it.

Because a locust antenna responds to many different odors, the same robot could in principle be retrained to hunt for new targets without training a new dog. It can also take a reading in seconds. The authors estimate more than 2,000 measurements a day, compared with the 20 to 100 minutes a standard lab instrument needs per sample.