ENabling Self-Driving in Uncertain Real Environments

Introduction

ENSURE addresses the challenge of self-driving in uncertain situations of the real world. Due to the difficulty of reasoning in complex real-world scenarios, self-driving remains one of the most difficult research problems today.

Safety and Navigation

For safe navigation, the driving agent needs to be able to anticipate the consequences of its actions. Current solutions are reactive without any planning for what might happen in the future. This poses major safety issues and delays the deployment of self-driving vehicles.

Risks of Inaction

Without a change in our approach to self-driving, we risk not only realizing fully autonomous driving but also half-baked solutions that endanger lives in uncertain situations. The future is inherently uncertain due to some scene structures such as intersections and the unknown intentions of the other agents.

Sources of Uncertainty

The errors in the perception of the scene and the prediction of the future cause another type of uncertainty. Furthermore, there are rarely encountered situations that might require passing the control to the human driver, such as an unknown object on the road.

Proposed Solution

As a way of managing uncertainties in the real world, ENSURE proposes a world model to predict the future with different types of uncertainty in a compact bird's eye view representation.

Implementation Strategy

To realize the potential of the world model, ENSURE will put it into action first online in simulation and push its performance to the limit under a controlled setting.

Ambitious Goals

The most ambitious goal of ENSURE is to learn to drive in an offline manner from already collected real driving data based on the predictions of the world model. The different types of uncertainties will be used to safeguard against the model's expected failures in the offline setting.

Conclusion

Every step of ENSURE will build towards enabling end-to-end driving in the real world, and its success in achieving this goal will allow similar success stories in other domains that require reasoning under uncertainty.