go back

Risk & Planning

Intention Estimation using Predictive Risk Maps

A key attribute of intelligent mobile systems is the capability to smoothly navigate in highly dynamic environments. This involves the evaluation of own behavior alternatives in the context of the anticipated scene dynamics, i.e., the predicted evolution of the other scene elements. The associated risk can be represented in so-called “risk-maps”, which quantify risk as a function of prediction time and ego-behavior.

Two factors influence the behavior planning: (1) the risk associated with a determined behavior and (2) resulting utility or gain in terms of mobility goals and comfort constraints. The optimal behavior is a tradeoff between risk and utility given by a path through the risk map which avoids the peaks of high risk. Furthermore, the risk map representation can also be used as a model for predicting the intended behavior of others, and for detecting whether they follow a risk-minimizing strategy or not.

Risk & Planning

Planing complex behaviors in complex environments means taking risks. A risk is the combination of the severity and the probability of a future aversive event. We might feel uncomfortable that artificial intelligent systems take risks. However, we already now rely on technology when we enter airplanes or stand on bridges. We constantly take risks, but we have learned to relate risk to utility and to social norms. Complex planning in systems means finding the right trade-off between utility and risk under the premise of machine ethics.

  • How do we move smoothly through a crowded airport hall?
  • How do we plan an overtaking manoeuvre on a busy highway?

The challenge when estimating risks is the uncertainty related to the prediction of the future events due to the unknown future behavior of others.
We use a variety of risk types and different prototypical behavior alternatives to gain a risk landscape which maps the times and probabilities of the highest risks. Taking additionally movement constraints and mobility gain factors into account, a near-optimal behavior can be planned.

For more information

J. Eggert, S. Klingelschmitt and F. Damerow, “The foresighted driver: Future ADAS based on generalized predictive risk estimation”,
Proc. 3rd Int. Symp. FAST-zero, 2015, pp. 93-100.