Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision
Previous revision
Member:sungbeanJo_paper [2021/03/04 23:09]
sungbean
Member:sungbeanJo_paper [2021/04/21 22:08] (current)
sungbean
Line 1: Line 1:
-For all controllers,​ the observation o is the currently +get_config_param active timestamp_mode  
-observed image at 20088 pixel resolution. For the measurement + TIME_FROM_INTERNAL_OSC 
-m, we used the current speed of the car, if available (in +get_config_param active multipurpose_io_mode 
-the physical system the speed estimates were very noisy and + OUTPUT_OFF ​ 
-we refrained from using them). All networks are composed +get_config_param active sync_pulse_in_polarity 
-of modules with identical architectures (e.g., the ConvNet + ACTIVE_LOW 
-architecture is the same in all conditions). The differences are +get_config_param active nmea_in_polarity 
-in the configuration of modules and branches as can be seen + ACTIVE_HIGH 
-in Figure 3. The image module consists of 8 convolutional +get_config_param active nmea_baud_rate 
-and 2 fully connected layers. The convolution kernel size is 5 + BAUD_9600 
-in the first layer and 3 in the following layers. The first, third, +
-and fifth convolutional layers have a stride of 2. The number +
-of channels increases from 32 in the first convolutional layer +
-to 256 in the last. Fully-connected layers contain 512 units +
-each. All modules with the exception of the image module +
-are implemented as standard multilayer perceptrons. We +
-used ReLU nonlinearities after all hidden layers, performed +
-batch normalization after convolutional layers, applied 50% +
-dropout after fully-connected hidden layers, and used 20% +
-dropout after convolutional layers. +
-Actions are two-dimensional vectors that collate steering +
-angle and acceleration:​ a = hs; ai. Given a predicted action +
-a and a ground truth action agt, the per-sample loss function +
-is defined as +
-All models were trained using the Adam solver [16] with +
-minibatches of 120 samples and an initial learning rate of +
-0:0002. For the command-conditional models, minibatches +
-were constructed to contain an equal number of samples with +
-each command.+
Navigation