Accelerating Autonomy in Space: From In-Orbit Demonstration to Standardised Deployment

A series of recent ESA-funded activities has advanced the state of the art in high-performance, energy-efficient on-board AI for space systems using Klepsydra software. These activities collectively demonstrate the applicability of lightweight, pipelined inference frameworks across a wide spectrum of space-qualified and radiation-hardened processing architectures.

The first activity achieved a successful in-orbit demonstration of Klepsydra AI inference engine on an operational satellite equipped with an NVIDIA Jetson TX2i. The experiment confirmed the reliability and performance of the underlying lock-free data-pipelining architecture in a real operational context, complemented by extended functionality tests on a flatsat platform.

A second activity investigated the deployment of AI algorithms on a complex radiation-hardened co-processor (HPDP-40), using convex optimisation methods to distribute computational workloads efficiently without requiring low-level hardware-specific coding. This work illustrates how advanced optimisation techniques can support autonomous and high-availability space missions on specialised European accelerators.

The final ESA activity focused on porting and optimising the same AI framework for GR740 and GR765 space-grade processors, covering LEON4/LEON5 and RISC-V NOEL-V architectures under both Linux and RTEMS. The study introduced convex-optimisation-based multicore tuning and initiated the first steps toward compliance with ECSS-E-40 and ECSS-Q-80 software standards.

In parallel, several internally funded investigations were conducted with industrial and research partners, including support for the HPSC architecture (in collaboration with Microchip), integration of VxWorks on the Teledyne e2v LX2160 (together with WindRiver), and work on enhancing cybersecurity protection mechanisms for on-board applications. These complementary efforts broaden the applicability of our AI framework to next-generation flight processors and safety-critical software environments.

Taken together, the ESA-funded and internally funded activities demonstrate the feasibility of deploying efficient, portable AI applications across heterogeneous space hardware platforms. They also illustrate a systematic progression toward flight-proven, radiation-tolerant, and ultimately space-qualified AI software suitable for future autonomous and data-intensive space missions.