Enhancing embedded performance and resource efficiency of the Mapless Autonomy Platform with the new 2024.01 version
driveblocks, a technology company providing perception software for mapless autonomous driving, rolls out an updated version of its core product, the Mapless Autonomy Platform with major improvements with respect to embedded deployments. The software toolkit leverages a combination of transformer neural networks and explainable sensor-fusion to provide vehicle OEMs and system integrators with a ready-to-use software library for various perception tasks. These include object detection, classification, drivable space detection, driving corridor / road model detection, as well as sensor-fusion and tracking. The software can be used for off-road applications, like agriculture, container terminals, or mining, as well as for highway driving.
Merging all perception tasks into a multi-task network
While the 2024.01 release is mainly oriented around performance improvements, it holds some major refactorings and feature additions:
Re-structuring of the core transformer neural networks to make heavy use of multi-task learning and inference. This makes it more data efficient during the training phase and ensures that multi-task inference is only slightly more resource intense than single-task inference.
Enhancing the resolution and recovery capabilities of the environment model and sensor-fusion in cases where sensor data provides conflicting inputs or sensor inputs have been occluded for a long time span.
Fixing of some rare error events when operating the software continuously for multiple hours.
Several minor improvements to decrease memory footprint and enhance compute efficiency.
driveblocks, a start-up based in Garching near Munich, offers a modular software platform for automated and autonomous driving in the commercial vehicle sector. The founding team looks back on many years of experience in the research and development of algorithms for automated driving and has already successfully participated in several research competitions, including the Indy Autonomous Challenge, in the USA.
The core technology is a combination of transformer neural networks and geometrically interpretable sensor-fusion to create a local environment model based on sensor data. In contrast to existing technologies based on so-called high-definition maps, this eliminates the need for costly pre-mapping and constant updating of these maps in the cloud. The offering is complemented by other compatible algorithms, providing customers with a full feature perception solution for off-road as well as highway applications.
The driveblocks software platform is sold both in individual modules and as a system solution via a licensing model and can be operated in various application areas, for example in mines, container terminals and on the highway.