Camera and control latencies are a crucial problem for teledriving, we need to understand which features are good predictors for understanding what causes latencies.

camera latency spike events

In this report we analyse:

• which features are important for latency prediction
• which features are important for spike prediction
• how early can we reliably forecast a spike event

We start from this previous analysis where 19 features were used to predict e2e_latency .
Data were collected upsampling all features up to a second, defining the spike as a latency lager than 400ms, filtering out the warehouses and extracting the features .

notebooks of the previous analysis

Features are resampled every 4 seconds, plotted , and a xgboost classifier is applied on the data.
The prediction is on 4 seconds bins where the maximum of e2e_latency is above 400ms is classified as label:1 .

Report on the data quality, ingestion, stream, plausibility

firing behaviour for telemetry features

Every sensor has a topic which fires to the backend at a rate from 1Hz to 50Hz. networking information comes every second while computing and vehicle_dynamics more frequently.
We spot which features are redundant and we check obvious dependecies with camera_latency

Preparation of the datasets for the training, feature selection and denoising

time series on deci seconds

We prepare different data sets:

• telemetry data, one second sampling
• telemetry data, one deci second sampling
• network_log data, one deci second sampling

For each we create a series of spike events (+/- 60 seconds from spike) and created a denoised average of all series

Investigation on time ordering of events, investigating causal dependecy

time series on deci seconds

We define a method to calculate the time persistance and the time shift between features.
Persistance shows more interesting results than correlation and really points out when two features have a causal relationship.
Time shift is a delicate measurement since doesn't always lead to consistent results. We show that networking features are mostly preceeding spikes and we compute the time shift for every single series to classify networking related spikes from computing related.

Prediction of camera_latency to spot the most relevant features for training the model.

spike prediction with networking features

We iterate over different models and feature sets to identity the most performant combination.
We predict as well the risk of a spike.
We end up using a long short term memory with few layers.

We use the prediction model to run a forecast on the camera_latency for the next few seconds.

forecast for every series, for every starting point

We calculate the accuracy of the forecast depending on the anticipation time. The reliability of the model increases close to the spike.

We investigate the spatial dependency of spikes.

latency per geohash

We clearly see a spatial pattern where the latency is larger in particular areas of the city.