IVNTR Light

A light weight implementation of IVNTR algorithm: paper, web. This repo includes:

• A simple Pick-Place gym environment (based on PRBench)
• A (batched) Task-then-motion planner that collects trajectories with operator/option annotations.
• Simple top-down effect-enumeration for predicate discovery on these trajectories.
• Task-then-motion planning with discovered predicates.

🔧 Installation

We strongly recommend uv. The steps below assume that you have uv installed. If you do not, just remove uv from the commands and the installation should still work.

# Install this repo
git clone https://github.com/Jaraxxus-Me/ivntr_light.git
cd ivntr_light
git submodule update --init

# Create venv
uv venv --python=3.11
source .venv/bin/activate
# Install skill_ref
uv pip install -e .[develop]
# Third-party dependencies
uv pip install -e third-party/prpl-mono/prpl-utils
uv pip install -e third-party/prpl-mono/prbench
uv pip install -e third-party/prpl-mono/relational-structs
uv pip install -e third-party/prpl-mono/toms-geoms-2d

🔬 Check Installation

1. Run ./run_ci_checks.sh. It should complete with all green successes.
2. Additionally, see videos for a simple dynamic2d execution video with given planner :)
3. Test pre-trained neural predicates:

   # Download data
   python3 scripts/download_data.py
   # Run test
   pytest -s -v tests/approaches/test_discovered_pred_skills.py::test_loading_learned_predicates_blocked_stacking
   

🔍 Guidelines for Understanding

Basics of Symbolic Planning

# Understanding PDDL's Matrix Formulation
pytest -s -v tests/utils/test_op_matrix_converting.py
# Understanding quantifiers on predicates
pytest -s -v tests/utils/test_quantified_predicates.py
# Task planning
pytest -s -v tests/utils/test_task_planning.py

Environment and Data Collection

# Understand the environment
pytest -s -v tests/benchmarks/test_blockedstacking_env.py
# How a task-then-motion planner works in (batched) gym.Env
pytest -s -v tests/approaches/test_pure_tamp.py
# How to collect trajectoris with the planner in (batched) gym.Env
pytest -s -v tests/datasets/test_collect.py

Operator and Predicate Learning

# How to learn operators with given predicates
pytest -s -v tests/approaches/test_op_learning.py
# How to learn predicates by enumerating their effects across operators (This takes a long time due to exhaustive enumeration, ~9 hours).
pytest -s -v tests/approaches/test_pred_learning_topdown.py
# How to learn predicates by bilevel learning the effects and classifiers (not tested here, in general should be more efficient).
pytest -s -v tests/approaches/test_pred_learning_bilevel.py

Planning with Discovered Predicates

# How to load discovered predicates/operators and bind skills to them.
pytest -s -v tests/approaches/test_discovered_pred_skills.py