Safety Chip

Codebase for ICRA paper: Plug in the Safety Chip: Enforcing Constraints for LLM-driven Robot Agents

Installation

1. Download Virtual Home simulator, and put them under virtualhome_v2.3.0/simulation.
2. Creating the environment by:

conda env create -f environment.yml

Prepare for Experiments

Set environment variable for OpenAI API key for current shell

export OPENAI_API_KEY=<YOUR_API_KEY>

Files

• src: all scripts for the project are stored here
  • exp_virtualhome_manip.py: script for virtualhome experiments
  • exp_spot.py: script for robot demo. Doesn't require virtualhome simulator installed
  • constraint_module.py: currently only used for translating natural language into LTL formulas
  • lang2ltl.py: library for enhanced version of lang2ltl framework. Can translate predifined predicates
  • precond.py: precondition module used for robot demo on baseline model and safety chip.
  • get_embed.py: encode action list or object list for grounding
  • utils.py
• results: experimental results. Two runs for robot demo. Two sets of results for virtualhome experiments: one with non-expert provided utterances, the other with expert provided utterances.
• virtualhome_v2.3.0/dataset: datasets for virtualhome exp and robot demo
• robot: nav graph scanned by Boston Dynamics Spot robot

Running Experiments

1. for virtualhome experiments, install virtaulhome simulator, export your OPENAI_API_KEY. Then you should be able to run exps by (an example)

python src/exp_virtualhome_manip.py --exp rooms --example_fname 0_1 --constraint_num 5 --safety_level full

If you place the embedding files somewhere, don't forget to specify them using the flags.

2. Similarly, you can also generate plans for robot demo by:

python src/exp_spot.py --example_fname task1 --constraint_num 10 --safety_level full