The Role of Predictive Uncertainty and Diversity in Embodied AI and Robot Learning

TL;DR

The paper surveys how predictive uncertainty and diversity influence embodied AI and robot learning, detailing the sources, types, metrics, and methods to quantify uncertainty, along with calibration techniques. It then shows how to leverage uncertainty across perception, mapping, planning, and control, including exploration and imagination through simulation and domain randomization. By contrasting evaluative and generative paradigms and emphasizing epistemic uncertainty, the work highlights pathways to improve robustness, generalization, and data efficiency in real-world robotics. The insights provide a practical blueprint for integrating uncertainty quantification, calibration, and diverse scenario generation into embodied AI pipelines to enhance safety, trust, and deployability.

Abstract

Uncertainty has long been a critical area of study in robotics, particularly when robots are equipped with analytical models. As we move towards the widespread use of deep neural networks in robots, which have demonstrated remarkable performance in research settings, understanding the nuances of uncertainty becomes crucial for their real-world deployment. This guide offers an overview of the importance of uncertainty and provides methods to quantify and evaluate it from an applications perspective.

Figures (9)

• Figure 1: Two paradigms of diversity. (a) In the evaluative paradigm, the machine learning model provides various hypothesis with associated likelihoods based on data it has seen. (b) In the generative paradigm, we need to generate hypothetical outcomes.
• Figure 2: An example of epistemic uncertainty. (a) A rover builds an elevation map of Mars martian to make decisions. If we have a distribution of maps instead of a single map, we can have different decision options, for instance to minimize risk, power consumption, etc. (b) The pits behind the boulders are not visible from the rover's front view. However, if the rover quantifies the epistemic uncertainty, then it knows how to make safe decisions. To represent this uncertainty, in its simplest form, we need the mean and variance of elevation senanayake2017bayesian.
• Figure 3: Sources of uncertainty.
• Figure 4: Types of uncertainty. Aleatoric uncertainty (known unknowns) is due to the inherent randomness of data whereas epistemic uncertainty (unknown unknowns) is due to lack of data. If we do not have data in a particular region of the input space or if we try to predict the future, the epistemic uncertainty is high.
• Figure 5: We need many models to quantify the epistemic uncertainty. This can be done explicitly by having multiple models as in ensembles or implicitly by introducing probability distributions over the parameters (weights of the neural network) of the ML model.