Proactive MDP-Based Collision Avoidance Algorithm for Autonomous Car

Distributed Autonomous System Laboratory

Collaborators and Students: 

Denis Osipychev, Duy Tran

Dr. Weihua Sheng

Project Description: 

Avoiding collisions with other vehicles on the highway require autonomously driven cars to make decisions using a non-stationary MDP and the Markov assumption.

The particular case considers the simulation of a three-lane highway with a robotic car approaching an intersection in presence of human drivers. The intentions of the human drivers are recognized by using hidden Markov Models from the car's data and the driver's facial expression. This intention is used as a switching means for the behavior models trained from the observation of all other drivers (now it considers 4 intentions – keep going, change lane left/right, slow down). This modeled behavior is a stochastic transition matrix in the grid world and used to build a reward model in the location/time domain. Then, the policy iteration is accomplished, finding the optimal policy of actions that reduce the chance of visiting the exact same location/time states as other cars do; avoiding collision.


Video: Speed up and go around Video: Slow down and veer
Collision Avoidance

More details: Denis Ospiychev, Duy Tran, Weihua Sheng, Girish Chowdhary, Proactive MDP-Based Collision Avoidance Algorithm for Autonomous Car, NIPS 2014, Autonomous Learning Robots workshop


DASLab supports the RLPy project