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 22:29]
sungbean
Member:sungbeanJo_paper [2021/04/21 22:08] (current)
sungbean
Line 1: Line 1:
-Assume that each observation o = hi;mi comprises an +get_config_param active timestamp_mode  
-image i and a low-dimensional vector m that we refer to as + TIME_FROM_INTERNAL_OSC 
-measurements,​ following Dosovitskiy and Koltun [9]. The +get_config_param active multipurpose_io_mode 
-controller F is represented by a deep network. The network + OUTPUT_OFF ​ 
-takes the image i, the measurements m, and the command +get_config_param active sync_pulse_in_polarity 
-c as inputs, and produces an action a as its output. The + ACTIVE_LOW 
-action space can be discrete, continuous, or a hybrid of these. +get_config_param active nmea_in_polarity 
-In our driving experiments,​ the action space is continuous + ACTIVE_HIGH 
-and two-dimensional:​ steering angle and acceleration. The +get_config_param active nmea_baud_rate 
-acceleration can be negative, which corresponds to braking or + BAUD_9600
-driving backwards. The command c is a categorical variable +
-represented by a one-hot vector. +
-We study two approaches to incorporating the command +
-c into the network. The first architecture is illustrated in +
-Figure 3(a). The network takes the command as an input, +
-alongside the image and the measurements. These three +
-inputs are processed independently by three modules: an +
-image module I(i), a measurement module M(m), and a +
-command module C(c). The image module is implemented +
-as a convolutional network, the other two modules as fullyconnected +
-networks. The outputs of these modules are concatenated +
-into a joint representation:​.+
  
-The control module, implemented as a fully-connected network, 
-takes this joint representation and outputs an action 
-A(j). We refer to this architecture as command input. 
-It is applicable to both continuous and discrete commands 
-of arbitrary dimensionality. However, the network is not 
-forced to take the commands into account, which can lead 
-to suboptimal performance in practice. 
-We therefore designed an alternative architecture,​ shown in 
-Figure 3(b). The image and measurement modules are as described 
-above, but the command module is removed. Instead, 
-we assume a discrete set of commands C = fc0; : : : ; cKg 
-(including a default command c0 corresponding to no specific 
-command given) and introduce a specialist branch Ai 
-for each of the commands ci. The command c acts as a 
-switch that selects which branch is used at any given time. 
-The output of the network is thus 
-We refer to this architecture as branched. The branches Ai 
-are forced to learn sub-policies that correspond to different 
-commands. In a driving scenario, one module might specialize 
-in lane following, another in right turns, and a third in 
-left turns. All modules share the perception stream. 
Navigation