basselfa / arena-bench-nav-prediction Goto Github PK

a) How many people
b) How to increase the speed (Rike)
c) Data recording on cloud

Only DWA and Burger

Create a table and analyze every data metric. What could go wrong, how it could interfere. What would collide a lot in all phases. Meeting Sat 10.12. at 12:00

1. tart collecting other data from other robots (same planner)
2. Create one DNN that takes data from multiple robots and outputs
3. Decide if one DNN for all output parameters is better or each output parameter should have their own DNN

Check how DNN with only one output parameter can be combined into one big DNN.
Make sure that output parameters add up to 1

Integrate the data recorder and pipeline into the new arena

Make result good for paper

Meet with Bassili, decide on how data should look like, analyze old data to see which input features are relevant

Fix Bug
Evaluate realistic map settings
Evaluate necessary dataset compositions (How many runs per map setting)
Set up plan for data recording
Create different datasets (different planners, Ni Lous maps,...)
Test different robots and planners
Check the tradeoff of complexity vs variability in the dataset

Check interesting papers for NN architecture

evaluate the robot performance on complicated maps (maps which have low robot performances)

Create different datasets (different planners, different robots, Ni Lous maps)

a) State of the art of navigation performance

Extract input parameters that had an significant influence on the output parameters

Success_rate, Collision_rate and Timeout_rate are not adding up to 1. Check out why and fix it

Check the tradeoff of complexity vs variability in the dataset

a) Current datasets for navigation performance
b) Additional recording metrics
c) Data distribution in dataset

Analyze related work section of project paper

The World complexity calculation sometimes returns Null values for distance(normalized), distance.avg, distance.variance. Check what causes this and solve the problem.

Actually not necessary, but will lead to more readable data.

Create one DNN with one planner and one robot which outputs all performance metrics

Check which map settings create impossible maps and decide how the setting has to be for a realistic map to exist

Planners that work:

• teb
• dwa
• mpc

Planners that don't work:

• arena
• rosnav
• rlca

Create the base architecture for a NN for navigation prediction from which other NN can be derived of.
Analyze which input variables are inportant from NN view

This occurs sometimes when running the pipeline. It can happen with the MainRecorder and also with the RobotRecorder. When it does, the node crashes so the corresponding data is not recorded.

Reason:

• Maybe it has to do with the topic sending data to the recorder before the current_time attribute is created in the recorder.

Possible Approaches:

• Add a waiting time to the other nodes so the recorder has enough time to create the attribute.
• Create the attribute before creating the callback function for the topic that sends the data.

Check how to create complicated maps on which will result in low robot performance metrics