Ajna: Generalized deep uncertainty for minimal perception on parsimonious robots.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Robots operate in dynamic environments with noisy sensors, leading to unreliable predictions.
• Neural networks excel at perception-driven predictions but often lack uncertainty quantification.
• Existing fusion methods require multiple observations, increasing computational load.

Purpose of the Study:

• To develop a mathematical formulation for heteroscedastic aleatoric uncertainty in neural network predictions.
• To introduce the Ajna network class for efficient uncertainty estimation in real-time robotic applications.
• To demonstrate the utility of uncertainty information in solving common robotics tasks without depth sensing.

Main Methods:

• Formulated a method to obtain heteroscedastic aleatoric uncertainty for arbitrary distributions.
• Developed the Ajna network, requiring minimal computation and a minor loss function modification.
• Utilized uncertainty from optical flow for obstacle avoidance, navigation, gap traversal, and object segmentation.

Main Results:

• The Ajna network enables real-time uncertainty estimation on resource-constrained robots.
• Demonstrated comparable performance to depth-based methods in obstacle dodging, cluttered navigation, gap flying, and object pile segmentation.
• Successfully evaluated on four common robotics and computer vision tasks using monocular vision.

Conclusions:

• The proposed Ajna network provides a generalized deep uncertainty method for robotics.
• Uncertainty cues from optical flow can effectively replace depth information for several robotic tasks.
• This approach enhances robot reliability and autonomy in complex, dynamic environments.