High-energy transient astrophysical phenomena, such as supernovae and binary neutron star mergers, benefit from a multi-wavelength investigation in which a space- or balloon-based omnidirectional telescope detects and localizes early high-energy emissions (such as a gamma-ray burst), then alerts a narrow-field follow-up instrument to observe the source. The high-energy telescope must provide a map that assigns to each sky location a likelihood that the source appears there. To issue prompt alerts despite limits on communication bandwidth and latency, it is desirable to compute this map aboard the high-energy telescope, but doing so requires rapid response while computing under stringent size, weight, and power constraints.

This work describes a real-time likelihood mapping implementation for Compton telescopes that is suitable for on-board computation. We use an adaptive multi-resolution approach and exploit parallelism and data reduction opportunities to achieve sub-second construction of high-resolution maps (HEALPix $N_{side}$=64) using a detailed instrument response matrix on a low-power ($<$ 10 W) embedded computing platform. We validate the speed and accuracy of our mapping approach on simulated high-energy transients from the third COSI Data Challenge.