Geisterhand

• Robotics at Compile Time
  Gneiss is meta-programmed using C++-23. Just provide a JSON config file describing your sensor setup and the compiler will strip away everything not needed from the header-only library. The result is a byte-perfect executable for your target hardware.
• Hard-Realtime Safe
  Gneiss does not use any sys-calls like dynamic memory allocation. Deployed on an embedded platform with a RTOS, it can guarantee a response time. To our knowledge, this is the only fixed-lag smoother capable of.
• Parallelization Everywhere
  Gneiss uses the new C++ execution model to generate abstract compute graphs. Compilers like gcc and AdaptiveCpp can generate parallelized code for multi-core processors, GPUs and many more.
• Probabilistic Accurate
  Gneiss is designed to give you meaningful variances and covariances. By modelling the movement as kinematic chain only relative poses with bounded covariances are estimated. Together with non-linear factor recovery a SLAM system can be designed superior to classical pose graph optimization.
• Robustness
  As front-end for camera streams gneiss uses Lucas-Kanade sparse optical flow tracking. This is very robust – even on terrain with self-repeating textures, like gravel or sand.
• Any Sensors
  For now, gneiss supports cameras and IMUs. Other sensors, like LIDAR, GNSS, 5G positioning, wheel odometry and velocity over ground sensors (for aerial and maritime applications) are coming soon.

We plan to setup a reinforcement pipeline that combines offline and online learning. That way, your robot can rapidly acquire a powerful world-model and then refine its policy by trial and error.

We plan to develop a SLAM back-end compatible to gneiss that uses non-linear factor recovery. By carefully modelling uncertainties your decision making and planning logic is precisely informed about the current state.

We plan to develop a physical A* search that can use maps and satellite images for efficient navigation and exploration into unknown terrain.