SYSTEM AND METHOD FOR RISK-AWARE BEHAVIORAL POLICY SELECTION BY AN AUTONOMOUS AGENT
A method for risk-aware policy assessment for an autonomous vehicle can include: collecting information associated with an environment of an ego vehicle; determining a set of policies; determining and assessing a set of risks encounterable (e.g., potentially encountered in the future) by the ego vehicle, and operating the ego vehicle based on the assessed risks. A system implementing the method can include a sensor suite, a computing system, a vehicle control system, and/or any other suitable set of components. In variants, the computing system can implement a multi-policy decision model, a risk model, an element selector, a policy generator, a fallback controller, policies (e.g., made up at least of policy elements, etc.), and/or any other suitable system components.
1 . A method for controlling a vehicle, comprising:
during a first time period:
sampling a first set of measurements of a set of environmental objects; and
based on the first set of measurements, determining a set of risks associated with the set of environmental objects; and
during a second time period:
based on the set of risks, determining a plurality of actions;
determining a first combination of actions from the plurality of actions;
determining a second combination of actions from the plurality of actions, wherein presence of at least one action differs between the first combination and the second combination;
determining a first set of vehicle control policies according to the first combination of actions;
determining a second set of vehicle control policies according to the second combination of actions;
sampling a second set of measurements of the set of environmental objects;
based on the second set of measurements:
performing a set of first forward simulations of the set of environmental objects according to the first set of vehicle control policies; and
performing a set of second forward simulations of the set of environmental objects according to the second set of vehicle control policies;
based on a comparison of the set of first forward simulations and the set of second forward simulations, selecting the first set of vehicle control policies; and
controlling the vehicle according to the first set of vehicle control policies.
2 . The method of claim 1 , wherein determining the set of risks comprises aggregating metrics from multiple simulations of the set of environmental objects, wherein the multiple simulations are performed during a same policy election cycle.
3 . The method of claim 1 , wherein the set of risks includes a first risk identified at a first timestep and a second risk identified at a second timestep distinct from the first timestep.
4 . The method of claim 1 , wherein the plurality of actions are determined based on a predetermined mapping between each risk and a respective set of actions.
5 . The method of claim 1 , wherein determining the first combination of actions comprises amending the first combination of actions such that the actions of the first combination are compatible.
6 . A method for controlling a vehicle, comprising:
determining a set of risks, each risk associated with at least one object of a set of objects in an environment of the vehicle;
based on the set of risks, selecting a plurality of actions from a predetermined set of actions;
from the plurality of actions, selecting a first combination of actions and a second combination of actions, wherein the first combination of actions and second combination of actions are different;
based on the first combination of actions, constructing a first set of policies for controlling the vehicle;
simulating the first set of policies applied to the set of objects in the environment;
determining a first set of metrics associated with the first set of policies;
based on the second combination of actions, constructing a second set of policies for controlling the vehicle;
simulating the second set of policies applied to the set of objects in the environment;
determining a second set of metrics associated with the second set of policies;
based on a comparison of the first set of metrics and second set of metrics, selecting the first set of policies; and
controlling the vehicle according to the first set of policies.
7 . The method of claim 6 , wherein the set of risks are determined using a forward simulation of the vehicle implementing a set of policies determined at a prior timestep.
8 . The method of claim 7 , wherein the set of risks are determined by aggregating metrics from a plurality of distinct forward simulations of sets of policies.
9 . The method of claim 8 , wherein the first set of policies and second set of policies are determined at the prior timestep.
10 . The method of claim 7 , wherein the set of risks include risks identified at multiple different timesteps.
11 . The method of claim 6 , wherein the first combination of actions comprises a set of semantic action identifiers with a predetermined mapping to types of risks included in the set of risks, wherein selecting the first combination of actions comprises using the predetermined mapping.
12 . The method of claim 6 , wherein the first combination of actions is determined independently of a subset of risks within the set of risks.
13 . The method of claim 6 , wherein the first simulation comprises multiple iterations of simulation using the same first set of policies, wherein behavior of another agent in the environment differs between simulations of the multiple iterations of simulation.
14 . The method of claim 6 , wherein selecting the first combination of actions comprises determining a compatibility of actions within the first combination of actions.
15 . The method of claim 6 , wherein the first set of policies is determined independently of a subset of risks within the set of risks.
16 . The method of claim 6 , wherein determining the set of risks comprises estimating a kinetic energy associated with avoiding a collision.
17 . The method of claim 6 , wherein a first risk of the set of risks is a risk associated with a conflict zone in the environment, wherein the first risk is associated with:
a location of the conflict zone;
a future time of reaching the conflict zone;
an estimated probability of the risk; and
an estimated severity of the risk.
18 . The method of claim 17 , wherein selecting the first combination of actions comprises using information selected from a set consisting of:
the location of the conflict zone;
the future time of reaching the conflict zone;
the estimated probability of the risk; and
the estimated severity of the risk.
19 . The method of claim 6 , wherein the first set of policies comprises a first set of constraints distinct from a second set of constraints of the second set of policies.
20 . The method of claim 6 , wherein the first set of policies comprises a vehicle controller.