Aleatoric and Epistemic Uncertainty Estimation in Autonomous Off-Road Navigation
Voiceover
Presenter(s)
Talay Kondhorn
Abstract or Description
Autonomous off-road navigation is a challenging field in robotics and machine learning because of the novelty and randomness of terrains. Even with large training datasets, many perception models struggle to consistently create accurate traversability cost maps. With uncertainty estimation, which to our knowledge has never been integrated into off-road navigation, the robot can become more aware of its confidence in interacting with the environment, allowing it to not only choose the most optimal path but also one where the predictions are most likely to be correct. We captured two types of uncertainty in perceptions, including aleatoric and epistemic uncertainty. Aleatoric is uncertainty inherent within the data. For example, we could never know for sure how deep a puddle is no matter the amount of training data given. Epistemic is uncertainty from the model. A model would have higher epistemic uncertainty on gravel if it was only trained on grass. We captured aleatoric uncertainty by splitting the deep neural network’s final layer into two heads, predicting variance along with the cost. During inference, the variance was higher in areas of tall grass, puddles, and bushes, confirming what we expected. Epistemic uncertainty was estimated by training an ensemble of models with different parameter initializations, predicting 23% higher uncertainty values in out-of-distribution datasets compared to in-distribution datasets, proving our model is robust. Future works can utilize aleatoric uncertainty to plan safer paths and epistemic uncertainty for uncertainty-aware exploration, driving the robot towards unfamiliar terrains to learn from a more diverse dataset.
Mentor
Wenshan Wang
Comments
Jess Kaminsky2 years ago
Excellent job defining Aleatoric Uncertainty and Epistemic Uncertainty. I'm not sure the visuals were particularly helpful in these descriptions but your verbal description was great. I do think there were a few terms that could have been helpful to define (model, output, uncertainty); my understanding is that these are very basic terms in the field but they might still be misunderstood by non-experts. I really liked your opening with the autonomous car mistake story and I wonder if you could have carried this example throughout your presentation. Your methodology and findings were fairly advanced for a non-expert but overall I got the sense that you were trying to make your content understandable and I appreciated that you continued to reference the two types of uncertainty as core to your work (if I'm a non-expert who just learned these terms, I was able to apply them several times therefore feeling more confident in my understanding). Additionally, your cadence of delivery was great, and that helped to digest new information.
•
Ken Hovis2 years ago
Great job of utilizing a story at the beginning to hook your audience and help them to understand the scientific exigence you are addressing with your research. You also did an excellent job of helping the viewer understand the difference between these two types of uncertainties and why they are important. It would have been helpful to use your cursor to highlight your visual as you were using it to help your audience follow and understand your findings better. I also would have liked to see you discuss more how these findings fit and contribute more broadly to the topic of autonomous vehicles.
•