US 11,938,957 B2
Driving scenario sampling for training/tuning machine learning models for vehicles
Eric Wolff, Boston, MA (US)
Assigned to Motional AD LLC, Boston, MA (US)
Filed by Motional AD LLC, Boston, MA (US)
Filed on Aug. 24, 2020, as Appl. No. 17/001,616.
Prior Publication US 2022/0055640 A1, Feb. 24, 2022
Int. Cl. B60W 50/08 (2020.01); B60W 50/00 (2006.01); G06N 20/00 (2019.01)
CPC B60W 50/085 (2013.01) [G06N 20/00 (2019.01); B60W 2050/0018 (2013.01); B60W 2050/0082 (2013.01)] 19 Claims
OG exemplary drawing
 
1. A method comprising:
assigning, using at least one processor, a set of initial physical states to a set of objects for a set of simulated driving scenarios;
generating, using the at least one processor, the set of simulated driving scenarios using the initial physical states of the objects in the set of objects;
determining, using the at least one processor, one or more failed driving scenarios in the set of simulated driving scenarios;
excluding, using the at least one processor, the one or more failed driving scenarios from the set of simulated driving scenarios, wherein the one or more failed driving scenarios comprise a driving scenario where acceleration or deceleration values of one or more objects in the set of objects are higher or lower than one or more specified threshold values;
training, using the at least one processor, a machine learning model using first samples of the set of simulated driving scenarios;
obtaining, using the at least one processor, tuning data comprising second samples of the set of simulated driving scenarios;
tuning, using the at least one processor, the trained machine learning model with the tuning data;
optimizing, using the at least one processor, a loss function of predictions output by the tuned machine learning model; and
causing, using the at least one processor, a vehicle to operate in an environment using the tuned machine learning model.