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 16:33]
sungbean 		Member:sungbeanJo_paper [2021/04/21 22:08] (current)
sungbean
Line 1: Line 1:
-Despite its positive performance,​ we identify limitations +get_config_param active timestamp_mode  
-that prevent behavior cloning from successfully graduating + TIME_FROM_INTERNAL_OSC 
-to real-world applications. First, although generalization +get_config_param active multipurpose_io_mode 
-performance should scale with training data, generalizing + OUTPUT_OFF ​ 
-to complex conditions is still an open problem with a lot +get_config_param active sync_pulse_in_polarity 
-of room for improvement. In particular, we show that no + ACTIVE_LOW 
-approach reliably handles dense traffic scenes with many +get_config_param active nmea_in_polarity 
-dynamic agents. Second, we report generalization issues + ACTIVE_HIGH 
-due to dataset biases and the lack of a causal model. We +get_config_param active nmea_baud_rate 
-indeed observe diminishing returns after a certain amount + BAUD_9600 
-of demonstrations,​ and even characterize a degradation of +
-performance on unseen environments. Third, we observe a +
-significant variability in generalization performance when +
-varying the initialization or the training sample order, similar +
-to on-policy RL issues [19]. We conduct experiments +
-estimating the impact of ImageNet pre-training and show +
-that it is not able to fully reduce the variance. This suggests +
-the order of training samples matters for off-policy Imitation +
-Learning, similar to the on-policy case [46].+

Trace:

Navigation