Cure

Causal Understanding and Remediation for Enhancing Robot Performance

Requirements

• Ubuntu 20.04 LTS
• ROS Noetic

Build Cure

1. Install Reval for performance measurement
2. Please use the following commands to build cure from source

git clone https://github.com/softsys4ai/cure.git
cd ~/cure && pip install -r requirements.txt
source /opt/ros/noetic/setup.bash
catkin build
source devel/setup.bash

N.b. If you face Catkin command not found, install sudo apt-get install python3-catkin-tools OR you can use catkin_make. For completeness we included the Husky and Turtlebot 3 source files in this repo to avoid complexity.

How to use Cure

CURE is used for tasks such as performance optimization and performance debugging in robotic systems. Given the cost and human involvement associated with collecting training data from physical robots for these tasks, CURE addresses these challenges by learning the performance behavior of the robot in simulations and transferring the acquired knowledge to physical robots. CURE also works with data from physical robots, we use simulator data to evaluate the transferability of the causal model.

• In simulation mode, CURE is compatible with any device utilizing Husky and Turtlebot in Gazebo environment.
• In reality mode, the performance measurements are directly taken from the physical robot. In the experiments, we have used Turtlebot 3 platform.
• Debugging mode, users can query the root cause of a certain functional and non-functional faults. For example, what is the root cause of the task failure and higher energy consumption

Arguments

optional arguments:
  -h, --help            show this help message and exit
  --robot {Husky_sim,Turtlebot3_sim,Turtlebot3_phy}
                        robotic platform
  --train_data          path of the training data (.csv)
  -root_cause           diagnose root-causes
  --outlier_data        path of the anomaly data (.csv)
  --f  [ ...]           functional properties (fixed: Task_success_rate)
  --nf  [ ...]          non-functional properties (default: [Energy,
                        Positional_error])
  --top_k               number of root-causes based on average causal effect
                        (default: 5)
  -l                    diagnose root-causes from the saved model [must be True
                        to use --model]
  --model               saved model path (.model)
  -opt                  perform optimization [must be True to use --budget,
                        --f1, --f2, --f1_pref, --f2_pref, --sc]
  --budget              number of evaluation to be performed (default: 10)
  --f1                  objectiive 1 (default: Energy)
  --f2                  objectiive 2 (default: Positional_error)
  --f1_pref             target f1 performance (default: Energy<=25.0)
  --f2_pref             target f2 performance (default: Positional_error<=0.1)
  --sc                  safety constraint, maintained distance from obstacle
                        (default: Obstacle_distance>=0.25)
  --tcr                 task completion rate (targets reached/total targets)
                        constraint (default: Task_completion_rate>=0.8)
  --hv_ref_f1           hypervolum reference for f1
  --hv_ref_f2           hypervolum reference for f2
  --init_trails         number of initial trails before fitting GP
  -viz                  show gazebo and rviz
  -l_opt                reload optimization from JSON
  --snap                optmization iteration to load from
  -v, --verbose         show statements during learning, effect estimation,
                        and optimization

Learning causal model

The causal model was trained on 1000 observational data obtained using Reval. The following commands can be used for training and saving the causal model. The saved model can later be utilized for both inference purposes and transferring knowledge. We have already included the saved models both for Husky and Turtlebot 3 in the cure/model directory.

python run_cure_MOO.py --robot Turtlebot3_sim --train_data data/obs_data/turtlebot3_1000.csv 

Multi-objective debugging

To find the root causes functional and non-functional faults from the saved causal model, please use the following command:

python run_cure_MOO.py --robot Turtlebot3_sim -root_cause --outlier_data data/bug/turtlebot3_outlier.csv -l --model model/care_Turtlebot3_sim.model --f Task_success_rate --nf Energy Positional_error Obstacle_distance --top_k 5

Multi-objective optimization

To run multi-objective optimization using cure, please run the following command:

python run_cure_MOO.py --robot Turtlebot3_sim --outlier_data data/bug/turtlebot3_outlier.csv -l --model model/care_Turtlebot3_sim.model --f Task_success_rate --nf Energy Positional_error Obstacle_distance --top_k 5 -opt --f1 Energy --f2 Positional_error --f1_pref 1.2 --f2_pref 0.15 --sc 0.25 --tcr 0.8 --hv_ref_f1 19.98 --hv_ref_f2 3.0 --budget 200

In this example, we used Energy and Positional_error as our two objectives, and Task_success_rate and Obstacle_distance as constraints.

Transferibility

Hardware

• Turtlebot 3 burger

Setup

1. Install the required dependencies for Turtlebot 3
2. Run roscore on remote PC
3. Run Bringup on Turtlebot 3 SBC

ssh ubuntu@{IP_ADDRESS_OF_RASPBERRY_PI}
roslaunch turtlebot3_bringup turtlebot3_robot.launch

4. Set Turtlebot 3 model on remote PC

$ export TURTLEBOT3_MODEL=burger

5. Run Cure To transfer the causal model learned from simulation to a physical robot (Turtlebot 3 sim -> Turtlebot 3 real), please use the following command.

python run_cure_MOO.py --robot Turtlebot3_phy --outlier_data data/bug/turtlebot3_outlier.csv -l --model model/care_Turtlebot3_sim.model --f Task_success_rate --nf Energy Positional_error Obstacle_distance --top_k 5 -opt --f1 Energy --f2 Positional_error --f1_pref 1.2 --f2_pref 0.18 --sc 0.25 --tcr 0.8 --hv_ref_f1 19.98 --hv_ref_f2 3 --budget 50

Real time results

To visualize the results in real time, please run the following commands in a separate terminal:

python live_plot.py --hv_ref_f1 19.98 --hv_ref_f2 3.2

More details abour Cure

For detailed instructions to reproduce our results, please use functionality and reproducibility