💻 computer science

Weakly-supervised Learning for Physics-informed Neural Motion Planning via Sparse Roadmap

This paper proposes Hierarchical Neural Time Fields (H-NTFields), a weakly-supervised framework that integrates sparse roadmap topological constraints with physics-informed PDE regularization to overcome the scalability and local minima limitations of existing methods in complex, high-dimensional motion planning.

Original authors: Ruiqi Ni, Yuchen Liu, Ahmed H. Qureshi

Published 2026-04-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ruiqi Ni, Yuchen Liu, Ahmed H. Qureshi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to walk through a messy, multi-room house to pick up a cup of coffee from the kitchen and bring it to the living room. The house is full of furniture, narrow hallways, and tricky corners. This is the "Motion Planning" problem.

For a long time, robots had two main ways to solve this, and both had big flaws:

The "Slow & Safe" Way (Classical Planners): Imagine a very cautious robot that stops at every single step to check if a chair is in the way. It draws a map of every possible path. It's very reliable and rarely crashes, but it's incredibly slow. By the time it figures out the path, your coffee is cold.
The "Fast & Risky" Way (Learning-Based AI): Imagine a robot that learns by watching thousands of videos of people walking. It gets very fast at guessing the path. But if it sees a room it hasn't seen before, or a chair in a weird spot, it might get confused, walk into a wall, or get stuck in a loop because it doesn't understand the physics of the space.

The Problem:
Existing "smart" robots that try to learn the rules of physics (like how waves spread out) often get lost in big, complex houses. They get stuck in local "dead ends" (like thinking the shortest way is through a wall) because they can't see the big picture of the whole house.

The Solution: H-NTFields (The "Hybrid GPS")
The authors of this paper created a new method called H-NTFields. Think of it as giving the robot a hybrid navigation system that combines a "rough sketch" with a "detailed GPS."

Here is how it works, using simple analogies:

1. The "Sparse Roadmap" (The Rough Sketch)

Imagine you are in a huge city. Before you start walking, you look at a very simple, low-resolution map that only shows the main highways and the big bridges between neighborhoods. It doesn't show every pothole or parked car, but it tells you: "To get from the Kitchen to the Living Room, you must go through the Hallway."

In the paper: This is the Sparse Roadmap. It's a simple graph of "safe zones" that gives the robot a rough idea of the global structure. It prevents the robot from getting lost in the big picture.

2. The "Physics Engine" (The Detailed GPS)

Now, imagine you have a high-tech GPS that knows the laws of physics. It knows that you can't walk through walls, and that you have to slow down when you are near a narrow doorway. It calculates the exact "travel time" to every point in the room.

In the paper: This is the PDE Regularization. It uses math (specifically the Eikonal equation, which describes how waves spread) to teach the robot how to move smoothly around obstacles.

3. The Magic Combination

The problem with just the "Rough Sketch" is that it's too vague (the robot might walk into a chair because the map didn't show it). The problem with just the "Detailed GPS" is that in a huge house, the math gets confused and the robot gives up or takes a wrong turn.

H-NTFields combines them:

The Rough Sketch acts as a "guardrail." It tells the robot, "Hey, the Living Room is over there, not through the wall." It keeps the robot on the right track globally.
The Detailed GPS fills in the gaps. It says, "Okay, you know you need to go to the Living Room, but here is exactly how to weave between the chairs and the table legs."

How They Trained It

Instead of needing a human to drive the robot around thousands of times (which is expensive and slow), they used a clever trick:

They built that simple "Rough Sketch" (Roadmap) automatically.
They took random points on that sketch and said, "What if the robot started here and wanted to go there?"
They used the Sketch to give the robot "weak hints" (e.g., "It takes at least 2 seconds, but no more than 5 seconds to get there").
They used the Physics Engine to make sure the robot didn't walk through walls.

This allowed the robot to learn fast and accurately without needing a massive library of human demonstrations.

The Results

When they tested this on a robot in a real kitchen (with a quadruped dog-robot carrying a can) and a robot arm in a messy cabinet:

Old AI methods: Got stuck, crashed, or failed to find a path in complex rooms.
Old Classical methods: Found a path but took way too long to calculate it.
H-NTFields: Found a smooth, safe path in a fraction of a second (0.15 seconds!) and successfully navigated the messy kitchen without dropping the can.

The Bottom Line

This paper is about teaching robots to be both fast and smart. By giving them a "rough map" to keep them oriented and a "physics brain" to handle the details, they can navigate complex, cluttered worlds (like our homes) much better than before, without needing to be taught by humans for every single new room.

1. Problem Statement

Robot motion planning involves finding collision-free paths in high-dimensional configuration spaces (C-spaces), particularly in cluttered, multi-room environments.

Limitations of Classical Methods: Sampling-based planners (e.g., RRT, PRM) offer theoretical guarantees but suffer from slow inference times and require high sampling density in complex spaces.
Limitations of Learning-Based Methods:
- Supervised methods rely on expensive expert demonstrations, which are hard to generate for high-DOF systems.
- Self-supervised Physics-Informed methods (e.g., Neural Time Fields or NTFields) solve the Eikonal equation to learn value functions without expert data. However, they struggle with scalability in complex environments. They often collapse into local minima, underestimate travel times over long horizons, and fail to maintain global consistency without dense sampling.

2. Methodology: Hierarchical Neural Time Fields (H-NTFields)

The authors propose H-NTFields, a weakly-supervised framework that unifies sparse topological supervision with physics-informed PDE regularization. The core insight is that sparse roadmaps provide global structural anchors, while PDE losses enforce local geometric fidelity.

A. Roadmap Construction for Weak Supervision

Instead of using a dense roadmap for direct planning, the authors construct a sparse roadmap to serve as a supervisory signal:

Free-Space Volume Sampling: Nodes are generated by sampling collision-free points and associating them with maximized "free spheres." New nodes are accepted only if they lie outside existing spheres, ensuring diversity and coverage without redundancy.
Sparse Connectivity: Nodes are connected via straight-line collision checks to form a graph capturing long-range connectivity (e.g., inter-room passages).
Start-Goal Perturbation: To generate training data, start and goal states are selected from roadmap nodes and perturbed within their associated free spheres. This exposes the network to local obstacle geometry and varied path lengths.
Weak Supervision Bounds: The roadmap provides upper and lower bounds on travel times ( $T_{ub}$ and $T_{lb}$ ) between perturbed start-goal pairs. These bounds act as weak labels, anchoring the neural field to feasible limits without requiring exact path labels.

B. PDE Regularization

To ensure the learned time field $T(q_s, q_g)$ respects the environment's geometry, the method employs a composite loss function derived from the Eikonal equation:

Eikonal Loss ( $L_E$ ): Enforces local consistency between the gradient of the time field and the inverse speed field (distance to obstacles).
Temporal-Difference Loss ( $L_{TD}$ ): Enforces the Bellman principle of optimality over finite time steps to prevent local overfitting and ensure value propagation.
Normal Alignment Loss ( $L_N$ ): Aligns predicted gradients with obstacle surface normals near boundaries.
Causality Weight ( $L_C$ ): A curriculum learning mechanism that prioritizes learning short paths before long ones to stabilize training.

C. Combined Training Objective

The final loss function combines PDE regularization with roadmap supervision:
$\mathcal{L} = (\lambda_E L_E + \lambda_{TD} L_{TD} + \lambda_N L_N + \lambda_R L_R) \cdot L_C$
Where $L_R$ is the roadmap-based loss penalizing predictions that fall outside the $[T_{lb}, T_{ub}]$ bounds. This synergy allows the PDE terms to refine local details while the roadmap anchors prevent global collapse.

D. Inference via Sampling-based MPC

Once trained, the value function $T(q_s, q_g)$ is embedded into a Sampling-based Model Predictive Control (MPC) framework. Candidate actions are sampled, scored by the predicted travel time, and a receding-horizon strategy selects the optimal trajectory. This approach mitigates sensitivity to local minima and supports multimodal path generation.

3. Key Contributions

Roadmap-Augmented PDE Framework: A novel hybrid training framework that unifies local geometric fidelity (via PDEs) with sparse global structure (via roadmaps), enabling scalable learning in complex environments.
Enhanced TD-NTFields: An extension of existing methods that integrates Eikonal, temporal-difference, and normal-alignment losses with roadmap-derived weak supervision, stabilized by a causality-based curriculum.
Robust Inference: Integration of the learned value function into a sampling-based MPC scheme, improving robustness and enabling real-time, continuous cost-to-go representation.
Empirical Validation: Demonstration of significantly higher success rates compared to PDE-only or roadmap-only baselines, scaling effectively to multi-room Gibson environments and high-DOF robots (6-DOF and 10-DOF).

4. Experimental Results

The method was evaluated on 18 Gibson environments (complex 3D indoor scenes) and real-world robotic platforms (UR5e arm and Unitree B1 quadruped).

Ablation Studies:
- Full Model (H-NTFields): Achieved 90.8% success rate (SR).
- PDE Only: SR dropped to 60.0% (failed on long-horizon tasks due to local minima).
- Roadmap Only: SR was 68.5% (lacked fine obstacle geometry).
- Conclusion: The combination is essential; PDEs provide smoothness, while roadmaps provide global connectivity.
Sample Efficiency:
- H-NTFields achieved 90.8% SR with only 50k training pairs.
- Prior PDE-only methods (TD-NTFields) required 200k pairs to reach only 76.8% SR.
- This demonstrates a 4x reduction in required data for superior performance.
Comparison with Baselines:
- vs. Classical Planners (RRT, PRM): H-NTFields is significantly faster (0.115s vs. ~1.0s) while maintaining comparable success rates (90.8% vs. ~99%).
- vs. Other Neural Methods: Outperformed NTFields (27.2% SR) and P-NTFields (37.2% SR) by a large margin.
- Real-World: Successfully executed long-horizon tasks (e.g., navigating a kitchen with a quadruped manipulator) in 0.15 seconds.

5. Significance

Bridging the Gap: H-NTFields successfully bridges the gap between neural field planners (fast, parallelizable, but brittle in complex spaces) and classical planners (robust, but slow and computationally heavy).
Scalability: By using sparse roadmaps as weak supervision, the method overcomes the "local minima" problem of pure PDE approaches without incurring the computational cost of dense roadmap generation or the data cost of expert demonstrations.
Practical Utility: The learned continuous value function serves as a reusable, amortized cost-to-go representation, making it highly suitable for integration into real-time control pipelines and MPC for complex, high-DOF robotic systems.