A probabilistic approach to planning and control in autonomous urban driving

Abstract

This paper considers the problem of decision mak ing and control for autonomous urban vehicles operating among other non-cooperating, possibly human controlled, vehicles. The difficulty in this problem stems from the fact that the behavior of the other vehicles is uncertain, and in many circumstances a collision cannot be prevented even under restrictive assumptions about the other drivers' actions. Tr aditional approaches that consider the worst-case actions of the other vehicles typically are inapplicable because the solutions are either too conserva tive or infeasible. This work proposes a new method for solving this problem using a chance constrained framework, which requires that the violation probability of all the constraints is guaranteed to be below a threshold. This formulation has the benefit of not only achieving traditional objectives such as minimization of fuel or travel time, but also hedges the nominal planned maneuver against potential crashes leading from the uncertainty in the other drivers' actions. The proposed framework is demonstrated on a passing maneuver example which exhibits a region where the passing vehicle cannot be guaranteed to prevent a collision in all circumstances.

Planning for cars that coordinate with people: leveraging effects on human actions for planning and active information gathering over human internal state

“…They are very capable when it comes to obstacle avoidance, lane keeping, localization, active steering and braking (Urmson et al 2008;Levinson et al 2011;Falcone et al , 2008Dissanayake et al 2001;Leonard et al 2008). But when it comes to other human drivers, they tend to rely on simplistic models: for example, assuming that other drivers will be bounded disturbances (Gray et al 2013;Raman et al 2015), they will keep moving at the same velocity (Vitus and Tomlin 2013;Luders et al 2010;Sadigh and Kapoor 2015), We equip autonomous cars with a model of how humans will react to the car's actions (a). We test the planner in user studies, where the car figures out that it can nudge into the human's lane to check their driving style (b, c): if it gets evidence that they are attentive it merges in front, expecting that the human will slow down; else, it retracts back to its lane (c).…”

Maximizing Road Capacity Using Cars that Influence People

“…where α i ∈ [0, 1] for all lanes and (6) constrains the solution to have overall autonomy level equal toᾱ, the autonomy level of the traffic feeding the road. To see this, observe that ∑ n i=1 α i c(α i ) is the sum of autonomous cars on all lanes and has to be equal toᾱ ∑ n i=1 c(α i ), which implies (6). Moreover, note that all the lanes have the same length, i.e.…”

Potential iLQR: A Potential-Minimizing Controller for Planning Multi-Agent Interactive Trajectories

“…The common approach to interactive trajectory planning is for a robot to make predictions of the future trajectories of other agents and plan reactively [7,8,9,10,11]. Planning reactively will make the agents decoupled and simplify the control problem.…”