CNR's Hybrid AI Policy Beats Infotaxis by 70% Tracking Moving Odor Sources
Researchers show that classic Infotaxis fails when odor sources move persistently, and a hybrid policy that predicts target trajectories fixes it.
ProTakeaways
- New paper reframes olfactory pursuit of moving sources as a POMDP with joint belief over position and velocity.
- Shows Infotaxis collapses to random-search performance when targets have high directional persistence.
- Proposes hybrid policy blending information gain with a greedy value function from a fully observable MDP.
- Hybrid achieves near-POMDP-optimal performance and outperforms Infotaxis by up to 70% at long persistence times.
- Robust under model mismatch, continuous run-and-tumble motion, and unsteady wake dynamics with about 25% gain.
- Full pseudo-code provided; extensions to online learning, turbulent flows, and adversarial prey are open.
Chasing something you can smell but can't see is a surprisingly hard control problem. Odor plumes are patchy, turbulence mixes them into misleading trails, and by the time a whiff reaches the searcher, the source has already moved on. A new paper from CNR, Tor Vergata, MaLGa Genoa, and ICTP tackles this directly, formalizing olfactory pursuit as a partially observable Markov decision process and showing that the field's default algorithm, Infotaxis, breaks down badly when the target actually moves.
The authors introduce a hybrid policy that mixes information gathering with predictive planning. It gets within striking distance of the quasi-optimal POMDP solution and beats Infotaxis by up to 70% in the hardest regime, while staying cheap enough to run in scenarios where the full POMDP is intractable.
Why Infotaxis stalls when prey run
Infotaxis, introduced by Vergassola, Villermaux and Shraiman in 2007, is the workhorse strategy for searching without gradients. The agent maintains a belief over source location and, at each step, moves to maximize the expected rate of information gain. It produces the zigzagging, casting behavior seen in moth flight and has been the reference algorithm for olfactory robotics for nearly two decades.
Infotaxis was designed for stationary targets. When the source moves with high directional persistence, holding a heading for many time steps, the agent keeps chasing information about where the target was, not where it is going. The new work makes this failure mode precise. A "blind spot" opens up where predator and prey move at comparable speeds; in that regime, pure exploration systematically fails to close the gap, and the algorithm degenerates toward a random search.