Map data compression methods implementing machine learning
Aspects of the present disclosure provide techniques for training a machine-learning model to compress map data for use online by an autonomous vehicle and techniques for compressing map data using the trained machine-learning model. A system includes a computing device configured to deploy a simulation environment initiating an instance of a virtual vehicle, execute iterations of a simulation of the virtual vehicle, wherein each iteration: deploys a set of map data compressed by the machine-learning compression model and causes the virtual vehicle to execute control operations based on the deployed set of map data, evaluate performance of the executed control operations by the virtual vehicle based on the compressed map data for each iteration, and train the machine-learning compression model to compress map data such that the evaluated performance of the executed control operations by the virtual vehicle exceeds a performance threshold.
1 . A system for training a machine-learning compression model to compress map data for use online by a vehicle, the system comprising:
a computing device configured to:
deploy a simulation environment initiating an instance of a virtual vehicle,
execute iterations of a simulation of the virtual vehicle, wherein each iteration: deploys a set of map data compressed by the machine-learning compression model and causes the virtual vehicle to execute control operations based on the deployed set of compressed map data, and wherein each iteration implements the machine-learning compression model that is iteratively evaluated and trained until the machine-learning compression model learns to compress the set of map data such that a performance of the virtual vehicle under simulation exceeds a performance threshold,
evaluate performance of the executed control operations by the virtual vehicle based on the compressed map data for each iteration, and
train the machine-learning compression model to compress map data such that the evaluated performance of the executed control operations by the virtual vehicle exceeds the performance threshold.
2 . The system of claim 1 , wherein the computing device is further configured to train the machine-learning compression model by:
identifying one or more proxy signals corresponding to aspects of the compressed map data based on the evaluated performance of the virtual vehicle, and
causing the training of the machine-learning compression model to preserve the one or more identified proxy signals when compressing map data.
3 . The system of claim 2 , wherein the one or more identified proxy signals comprises at least one of a smoothness of lane geometry, intersection positioning, or sign positioning.
4 . The system of claim 1 , wherein the computing device is further configured to:
execute iterations of the simulation of the virtual vehicle, wherein each subsequent iteration deploys an updated compressed map data based on training revisions made to the machine-learning compression model.
5 . The system of claim 1 , wherein the computing device is further configured to:
generate compressed map data with a trained machine-learning compression model; and
transmit the generated compressed map data to the vehicle.
6 . The system of claim 1 , wherein the instance of the virtual vehicle deployed in the simulation environment comprises at least one autonomous vehicle system.
7 . The system of claim 1 , wherein the instance of the virtual vehicle deployed in the simulation environment comprises at least one semi-autonomous vehicle system.
8 . The system of claim 1 , wherein the set of map data comprises geospatial data.
9 . A method for training a machine-learning compression model to compress map data for use online by a vehicle, the method comprising:
deploying, with a computing device, a simulation environment initiating an instance of a virtual vehicle;
executing, with the computing device, iterations of a simulation of the virtual vehicle, wherein each iteration: deploys a set of map data compressed by the machine-learning compression model and causes the virtual vehicle to execute control operations based on the deployed set of compressed map data, and wherein each iteration implements the machine-learning compression model that is iteratively evaluated and trained until the machine-learning compression model learns to compress the set of map data such that a performance of the virtual vehicle under simulation exceeds a performance threshold;
evaluating, with the computing device, performance of the executed control operations by the virtual vehicle based on the compressed map data for each iteration; and
training, with the computing device, the machine-learning compression model to compress map data such that the evaluated performance of the executed control operations by the virtual vehicle exceeds the performance threshold.
10 . The method of claim 9 , wherein training the machine-learning compression model further comprises:
identifying one or more proxy signals corresponding to aspects of the compressed map data based on the evaluated performance of the virtual vehicle; and
causing the training of the machine-learning compression model to preserve the one or more identified proxy signals when compressing map data.
11 . The method of claim 10 , wherein the one or more identified proxy signals comprises at least one of a smoothness of lane geometry, intersection positioning, or sign positioning.
12 . The method of claim 9 , wherein:
each subsequent iteration of the simulation deploys an updated compressed map data based on training revisions made to the machine-learning compression model.
13 . The method of claim 9 , further comprising:
generating compressed map data with a trained machine-learning compression model; and
transmitting the generated compressed map data to the vehicle.
14 . The method of claim 9 , wherein the instance of the virtual vehicle deployed in the simulation environment comprises at least one autonomous vehicle system.
15 . The method of claim 9 , wherein the instance of the virtual vehicle deployed in the simulation environment comprises at least one semi-autonomous vehicle system.
16 . The method of claim 9 , wherein the set of map data comprises geospatial data.