TL;DR
Flax is a new neuro-symbolic task planning framework that combines the speed of neural object pruning with the reliability of symbolic reasoning. By introducing a lightweight symbolic correction mechanism—solving a simplified, rule-relaxed version of the task to catch missing objects—Flax fixes neural mispredictions on the fly, boosting planning success rates by 20.82% while cutting down wall-clock planning time.
What Problem Does This Paper Address?
Robots operating in complex environments (like warehouses or cluttered rooms) must perform long-horizon task planning. Classical symbolic planners, such as Fast Downward, provide mathematical guarantees but suffer from a "combinatorial explosion" as the number of objects and potential actions grows.
To combat this, recent neuro-symbolic (NeSy) methods use Graph Neural Networks (GNNs) to predict which objects are "important" to the goal, pruning the rest to shrink the search space. However, these approaches have a critical vulnerability: if the neural network accidentally prunes even one essential object, the simplified task becomes completely unsolvable. The symbolic planner will then exhaust its entire time budget trying to solve an impossible problem, leading to catastrophic dead-ends. This paper addresses the lack of an efficient correction mechanism to detect and fix these neural pruning errors.
Methods
The authors introduce Flax (Neuro-Symbolic Relaxation), a three-step pipeline designed to keep the planning space compact yet safely solvable:
- Importance-guided Pruning: A GNN scores the importance of each object in the environment. Objects below a certain threshold are filtered out to create a bootstrap simplified task, which is handed to a symbolic planner.
- Rough-plan Enhancement (Symbolic Relaxation): If the planner fails or times out, Flax doesn't just lower the threshold blindly. Instead, it solves a rule-relaxed version of the full task. This relaxation simplifies the physics or constraints (e.g., ignoring certain types of obstacles), making it incredibly fast to solve. The resulting "rough plan" reveals which critical objects were mistakenly left out by the GNN, and those objects are immediately re-integrated.
- Complementary Rule Expansion: Finally, Flax applies lightweight, domain-specific symbolic rules to enforce local spatial and relational consistency (e.g., ensuring that if a specific location is included, the object occupying it is also included), producing a finalized, reliable task description.
Datasets
The framework was evaluated across three distinct domains:
- MazeNamo (Maze Navigation among movable obstacles): A newly introduced benchmark built in MiniGrid. It features grid sizes ranging from $10\times10$ to $15\times15$ and four difficulty levels (Easy, Medium, Hard, Expert), requiring robots to navigate, push heavy blocks, and stack light boxes.
- Isaac Sim: High-fidelity physics simulations mimicking a forklift operating in a cluttered warehouse environment.
- Real-World Hardware: Deployed directly on a Unitree Go2 quadruped robot equipped with a 6-DoF D1 manipulator and a LiDAR/camera setup.
Key Insights and Ideas
- Imperfection is Fine with a Safety Net: You don't need a flawless, massive neural network if you have a lightweight symbolic correction mechanism. By using a rule-relaxed symbolic planner as a validator, Flax catches neural blind spots faster than iterative neural threshold adjustments.
- Zero-Shot Scale Generalization: The GNN importance predictor was trained only on small $10\times10$ mazes but successfully generalized to massive $15\times15$ expert mazes. This demonstrates that learning relational object importance scales much better than learning direct end-to-end policies.
- Abstract to Physical Execution: The exact same model trained in abstract 2D simulations successfully drove a real-world quadruped robot to manipulate physical objects without any real-world retraining or fine-tuning.
Key Results
- Performance Boost: Compared to the state-of-the-art neuro-symbolic baseline (PLOI), Flax improved the average planning success rate by 20.82% and reduced the mean wall-clock planning time by 17.65%.
- Hard/Expert Dominance: On the most challenging $10\times10$ expert tasks, Flax achieved a +49.67% success rate improvement over PLOI. Classical planning alone completely failed (0% success rate) within the given time budgets on expert tasks across all map sizes.
- Sim-to-Real Success: In physical deployments, Flax achieved a 67% success rate across varied real-world layouts, with failures primarily caused by hardware/execution issues (like boxes slipping or rotating unexpectedly) rather than planning errors.
Limitations and Future Work
Currently, Flax relies on domain-specific relaxation rules that must be manually designed by human experts (e.g., deciding to remove light boxes to create the rough plan).
To overcome this, the authors point out several promising future directions:
- Automatic Rule Discovery: Developing methods to automatically learn or synthesize these symbolic relaxation rules from the PDDL domain definition itself.
- Online Refinement: Integrating an online learning loop so the neural network can update its importance predictor based on the mistakes caught by the symbolic relaxation step during deployment.
- Continuous TAMP: Extending the framework from discrete task planning to continuous, hybrid Task-and-Motion Planning (TAMP) environments.
Summary updated