Qiwei Du1, Bowen Li2, Yi Du1, Shaoshu Su1, Taimeng Fu1, Zitong Zhan1, Zhiepeng Zhao1, Chen Wang1
1: Spatial AI & Robotics (SAIR) Lab, Computer Science and Engineering, University at Buffalo
2: AirLab, Robotics Institute, Carnegie Mellon University
Accepted to IEEE Robotics and Automation Letters (RA-L), 2026
Real-world task planning requires long-horizon reasoning over large sets of objects with complex relationships and attributes, leading to a combinatorial explosion for classical symbolic planners. To prune the search space, recent methods prioritize searching on a simplified task only containing a few "important" objects predicted by a neural network. However, such a simple neuro-symbolic (NeSy) integration risks omitting critical objects and wasting resources on unsolvable simplified tasks. To enable Fast and reliable planning, we introduce a NeSy relaxation strategy (Flax), combining neural importance prediction with symbolic expansion. Specifically, we first learn a graph neural network to predict object importance to create a simplified task and solve it with a symbolic planner. Then, we solve a rule-relaxed task to obtain a quick rough plan, and reintegrate all referenced objects into the simplified task to recover any overlooked but essential elements. Finally, we apply complementary rules to refine the updated task, keeping it both reliable and compact. Extensive experiments are conducted on both synthetic and real-world maze navigation benchmarks where a robot must traverse through a maze and interact with movable obstacles. The results show that Flax boosts the average success rate by 20.82% and cuts mean wall-clock planning time by 17.65% compared with the state-of-the-art NeSy baseline. We expect that Flax offers a practical path toward fast, scalable, long-horizon task planning in complex environments.
Video:
We propose a task called Maze Navigation Among Movable Obstacles (MazeNamo). In a simplified MiniGrid environment, the scenario consists of a grid-based layout with three types of objects: walls (grey blocks), heavy boxes (blue blocks), and light boxes (yellow blocks). The agent can execute basic navigation actions, including moving forward and turning left or right, as well as three additional manipulative actions: push, pick up, and drop down. The following constraints govern object interactions:
- Heavy boxes can only be pushed.
- Light boxes can be both pushed and picked up.
- Light boxes can be dropped either on the floor or on heavy boxes but cannot be placed on other light boxes.
- At any given time, the agent can interact with only one box, meaning it cannot push or pick up multiple boxes simultaneously.
Create conda environment
conda create -n flax python=3.10 -y
conda activate flax
pip install -r requirements.txt
Add MazeNamo env into Minigrid and install Minigrid
bash scripts/create_mazenamo_env.sh
The dataset is in pddl_files/problems/mazenamo_problems.
Link the domain file and the training set to pddlgym:
mkdir pddlgym/pddl
ln -s $(pwd)/pddl_files/domains/mazenamo.pddl $(pwd)/pddlgym/pddl/mazenamo.pddl
ln -s $(pwd)/pddl_files/problems/mazenamo_problems/pddl_10x10_train $(pwd)/pddlgym/pddl/mazenamo
You can also generate new datasets using:
python src/generate_mazenamo_problems.py
where problem size and problem difficulty can be changed.
bash scripts/install_VAL.sh
Trained models are reused in batch scripts.
All experiments use models trained on 200 10x10 problems.
(If training hasn't been run before, weights will appear in model/. If model/ already exists, training will be skipped.)
# run experiments with one group of settings
bash scripts/run_mazenamo.sh
# run experiments with multiple groups of settings
bash scripts/batch_run_mazenamo.sh
# generate .gif for a specific problem or all the problems of a certain group of settings
bash scripts/batch_vis_mazenamo.sh
pip install isaacsim[all]==4.5.0 --extra-index-url https://pypi.nvidia.com
pip install usd-core # named pxr when importing, used to handle usd files
Download usd files here:
https://drive.google.com/file/d/1Jg4G-aZOua9H5BjGLgwRbY28F0TyK66f/view?usp=sharing
The file structure should be:
flax/
├── assets/
│ ├── Collected_Box_A09_40x30x23cm_PR_V_NVD_01/
│ ├── Collected_...
│ ├── 2x2_warehouse_shapes.npy
│ ├── 2x2_warehouse_base_0.usd
│ └── ...
└── ...
First run:
python src/run_isaacsim_gui.py
Then, choose a .usd file in the GUI.
For simplicity, we only used a 2x2 square warehouse as the base scene in our experiments (assets/2x2_warehouse_shapes.npy).
However, you can generate unique warehouse shapes using generate_random_warehouse() in src/generate_isaacsim_warehouse_shape.py, change the path of warehouse_shapes and run:
python src/generate_isaacsim_warehouse_base_scene.py
if you have customized requirements. You will get a .usd file of an empty warehouse (something like assets/2x2_warehouse_base_0.usd).
Generate from a grid map in pddl_files/problems/mazenamo_problems/:
python src/generate_isaacsim_mazenamo_from_gridmap.py
Generate from a handcrafted scene:
python src/generate_isaacsim_mazenamo_example.py
Remember to set the initial robot direction in solve_isaacsim_mazenamo_from_usd_forklift.py according to the output info of generate_isaacsim_mazenamo_from_gridmap.py.
python src/solve_isaacsim_mazenamo_from_usd_forklift.py
(Note: Rendering on the RTX 4090 GPU currently supports only 10×10 and 12×12 maps. Scaling up to a 15×15 map requires a 3×3 warehouse configuration; however, rendering the full scene sometimes exceeds the available GPU memory.)
We also introduced two challenging domains to demonstrate the generalizability of our Flax.
It builds on the classic Logistics benchmark, where trucks and airplanes transport packages across a network of cities. Unlike the original domain, our variant requires reasoning over an explicit connectivity graph of directed road and air links rather than relying on simplified city-level abstractions. It further increases difficulty by introducing unit-capacity vehicles, stackable packages, and locked road hubs that can only be unlocked via key packages at specific switch panels. These additions create tight resource constraints and long-range causal dependencies, making the planning task substantially more complex than standard Logistics.
mkdir pddlgym/pddl
ln -s $(pwd)/pddl_files/domains/difficultlogistics.pddl $(pwd)/pddlgym/pddl/difficultlogistics.pddl
ln -s $(pwd)/pddl_files/problems/difficultlogistics_problems/pddl_train $(pwd)/pddlgym/pddl/difficultlogistics
# run experiments with one group of settings
bash scripts/run_difficultlogistics.sh
python src/generate_difficultlogistics_problems.py
It is a challenging variant of the classic Sokoban puzzle, where an agent pushes boxes on a grid to designated goal cells without pulling or passing through obstacles. Unlike standard Sokoban, where all boxes are typically interchangeable, SokoMind Plus assigns specific boxes to individual goal locations while treating the remaining boxes purely as movable obstacles. This introduces tighter goal constraints and additional clutter, making long-horizon planning substantially harder than in the original Sokoban setting.
mkdir pddlgym/pddl
ln -s $(pwd)/pddl_files/domains/sokomindplus.pddl $(pwd)/pddlgym/pddl/sokomindplus.pddl
ln -s $(pwd)/pddl_files/problems/sokomindplus_problems/pddl_train $(pwd)/pddlgym/pddl/sokomindplus
# run experiments with one group of settings
bash scripts/run_sokomindplus.sh
python src/generate_sokomindplus_problems.py
Parts of this repository are derived from https://github.com/tomsilver/ploi and https://github.com/tomsilver/pddlgym. Many thanks to the original authors for sharing their code.
If you used our code in your research, or you find our work useful, please cite us as:
@article{du2025fast,
title={Fast Task Planning with Neuro-Symbolic Relaxation},
author={Du, Qiwei and Li, Bowen and Du, Yi and Su, Shaoshu and Fu, Taimeng and Zhan, Zitong and Zhao, Zhipeng and Wang, Chen},
journal={arXiv preprint arXiv:2507.15975},
year={2025}
}