Repository for "Cost-aware Discovery of Contextual Failures using Bayesian Active Learning", CoRL 2025
CARLA failure discovery for Mode 1 (misdetection due to bad visibility from distance) and Mode 2 (misdetection due to bad lighting)
This example uses a GiT+LLM for failure diagnosis as an expert Instructions:
- Clone the repo
- Create conda environment and execute the following script to check if things are working:
conda env create -f environment.yml
conda activate cfail
python3 scripts/carla/run_eci.py --seed SEED --num_init N1 --num_iter N2 --delta_light D1 --delta_dist D2 --radius R
Use SEED, N1,N2 to control the seed, number of iterations for initializing the prior using random sampling, number of iterations for Bayesian loop.
Control the severity of each failure mode using D1 and D2, (0 means low threshold on severity, 1 means high). We estimate severity in this example using the number of images in a simulation where misdetection happens. Min value should be 0.1, which corresponds to atleast one image with failure per simulation. R controls the sampling resolution, aka neighbourhood of each scenario, setting R higher will lead to broader sampling, low R will lead to finer sampling.