A Human-in-the-loop Confidence-Aware Failure Recovery Framework for Modular Robot Policies

Simulation framework implementation for the paper A Human-in-the-loop Confidence-Aware Failure Recovery Framework for Modular Robot Policies.

Project Website | Paper

Requirements

• Python 3.11+
• Tested on MacOS Catalina

Installation

1. Recommended: create and source a virtualenv.
2. pip install -e ".[develop]"
3. python -m amplpy.modules install coin -q

🚀 Quick Start

To generate the main paper plot (Fig. 4):

1. Running simulation experiments:

python experiments/run_experiment.py --variant all_variants

2. Create data directory:

mkdir -p experiments/results/[data_dir]
mv experiments/results/* experiments/results/[data_dir]

3. Create dataframes:

python experiments/pickles_to_df.py --data_dir experiments/results/[data_dir]

Note - you will need to run steps 1-3 3 times in total, one for each of the configurations in the main() method of experiments/run_experiment.py, where the required [data_dir] values are written in the comments.

Generate main paper plot (Fig. 4):

python experiments/plot_unified_grid.py --output_dir [output_dir]

To generate the appendix plots, please checkout the feature/noisy-experts branch.

📚 Citation

@inproceedings{banerjee2026modularhil,
      author    = {Banerjee, Rohan and Palempalli, Krishna and Yang, Bohan and Fang, Jiaying and Abdullah, Alif and Silver, Tom and Dean, Sarah and Bhattacharjee, Tapomayukh},
      title     = {A Human-in-the-Loop Confidence-Aware Failure Recovery Framework for Modular Robot Policies},
      booktitle = {Proceedings of the ACM/IEEE International Conference on Human-Robot Interaction (HRI)},
      year      = {2026},
}