Uploaded September 2023 | Updated September 2026, 3 weeks ago
We *open-source* the ORACLE library of methods on deep learned collision-free navigation of aerial robots that assume a) no access to a map of the environment or an estimate of the robot’s position, and presents robust sim2real transfer. ORACLE enables safe uncertainty-aware flight, while its visually-attentive variant (A-ORACLE) combines that capacity with implicit information sampling, and seVAE-ORACLE alters the architecture to offer modularization and partial training on both synthetic and real data (if available). For navigation, the method(s) only consider the current depth image (e.g., from an RGB-D or a passive stereo camera) and a partial state estimate (linear velocities, roll/pitch, angular rate around the z-axis), alongside a bearing vector to the desired direction. The two core navigation methods, ORACLE and seVAE-ORACLE focus either on a) uncertainty-awareness by considering the robot’s partial state estimate uncertainty and epistemic uncertainty over the neural net through a deep ensemble, or b) modularized encoding of the high-dimensional depth data through a semantically-enhanced variational autoencoder – allowing to retain information over hard-to-perceive obstacles in aggressively compressing latent spaces - before the collision prediction step. A-ORACLE uses also a mask over the depth image that highlights visually-attentive regions in order to select the action primitive that not only allows the robot to navigate the environment safely but also to attend (using its camera) to salient regions either from a bottom-up or a top-down perspective (depending how the mask is derived).
Open-Source code access: github.com/ntnu-arl/ORACLE
Detailed wiki: github.com/ntnu-arl/ORACLE/wiki
We *open-source* the ORACLE library of methods on deep learned collision-free navigation of aerial robots that assume a) no access to a map of the environment or an estimate of the robot’s position, and presents robust sim2real transfer. ORACLE enables safe uncertainty-aware flight, while its visually-attentive variant (A-ORACLE) combines that capacity with implicit information sampling, and seVAE-ORACLE alters the architecture to offer modularization and partial training on both synthetic and real data (if available). For navigation, the method(s) only consider the current depth image (e.g., from an RGB-D or a passive stereo camera) and a partial state estimate (linear velocities, roll/pitch, angular rate around the z-axis), alongside a bearing vector to the desired direction. The two core navigation methods, ORACLE and seVAE-ORACLE focus either on a) uncertainty-awareness by considering the robot’s partial state estimate uncertainty and epistemic uncertainty over the neural net through a deep ensemble, or b) modularized encoding of the high-dimensional depth data through a semantically-enhanced variational autoencoder – allowing to retain information over hard-to-perceive obstacles in aggressively compressing latent spaces - before the collision prediction step. A-ORACLE uses also a mask over the depth image that highlights visually-attentive regions in order to select the action primitive that not only allows the robot to navigate the environment safely but also to attend (using its camera) to salient regions either from a bottom-up or a top-down perspective (depending how the mask is derived).
Open-Source code access: github.com/ntnu-arl/ORACLE
Detailed wiki: github.com/ntnu-arl/ORACLE/wiki
