🤖 AI

SPADE: Sketch-guided Path Planning Augmented with Diffusion Experts

This paper introduces SPADE, a novel path planning framework for autonomous mobile robots that combines an enhanced ROS 2 annotation tool with a diffusion-based augmentation strategy to achieve superior generalization, robustness, and accuracy with significantly fewer trainable parameters than state-of-the-art methods.

Original authors: Charbel Abi Hana, Tatiana Ghantous, Mikael Khalil, Anthony Rizk

Published 2026-06-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Charbel Abi Hana, Tatiana Ghantous, Mikael Khalil, Anthony Rizk

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 teaching a robot how to drive a delivery truck through a busy warehouse. You want the robot to drive smoothly, avoid tight corners, and take specific routes that human managers prefer.

The Problem: The "Hard-to-Teach" Robot
Traditionally, teaching a robot these preferences is like trying to explain a complex dance move by writing a math equation. You have to program every single rule (e.g., "turn left if the distance is 5 meters"), which is incredibly difficult and often misses the "feel" of how a human actually drives.

Another method is Imitation Learning, where you let a human drive the robot (or draw a path on a map) and the robot tries to copy them. A previous system called SKIPP did this well: a human could draw a squiggly line on a map, and the robot would learn to follow that style. However, SKIPP had two big flaws:

The "Teacher" was weak: It struggled when the robot went into a new warehouse it had never seen before. It was like a student who memorized the answers to one specific test but failed the next one.
The "Notebook" was broken: The tool used to collect the human drawings was old, clunky, and hard to fix.

The Solution: SPADE (The "Master Chef" and the "Apprentice")
The authors created a new system called SPADE. They fixed the broken tool and introduced a clever training strategy that uses Diffusion Models.

To understand how SPADE works, imagine a Master Chef and a Young Apprentice:

The Master Chef (The Diffusion Model): This is a huge, incredibly smart, but very slow computer program. It has "tasted" thousands of perfect paths. It knows exactly what a smooth, safe, and efficient route looks like in any situation. However, it's too slow to run inside the robot while it's moving (it would take 45 seconds to plan a route that needs to happen instantly).
The Young Apprentice (The Lightweight Robot Model): This is the small, fast program that actually lives on the robot. It needs to make decisions in milliseconds.

How They Train Together:
Instead of just showing the Apprentice a picture of the path and saying "copy this," the SPADE system uses the Master Chef as a critic during training.

The Apprentice tries to draw a path.
The Master Chef looks at the Apprentice's drawing and says, "That's okay, but if I were drawing it, I would have curved it slightly more here to avoid a bump."
The Apprentice learns from this feedback, not just by copying, but by understanding the principles of a good path.

This allows the small, fast Apprentice to learn the deep wisdom of the slow, smart Master Chef without actually needing to be slow itself.

The New Tools
The team also built a brand-new, open-source "notebook" (a software tool based on ROS 2) that lets humans easily draw paths on 3D maps. This makes it much easier to gather high-quality examples for the robot to learn from, replacing the old, broken system.

The Results
When they tested this new system:

Better Accuracy: The robot's paths were much closer to the human's intended route (39% less error).
Better Generalization: The robot could handle new, unseen warehouses much better than before.
Efficiency: The "Apprentice" model is tiny. It uses 93.8% fewer computer resources (parameters) than the massive models usually required to get this level of performance. It's like getting the performance of a supercomputer using a standard smartphone processor.
Real-Time Speed: Because the final model is so small, the robot can think and move instantly, whereas the "Master Chef" alone would be too slow for real-world driving.

In Summary
SPADE is a way to teach robots to drive like humans by using a "slow but genius" AI to train a "fast but simple" AI. This results in a robot that is smart enough to handle new environments, fast enough to drive in real-time, and small enough to run on standard hardware, all while using a much better tool to collect the training data.

Technical Summary: SPADE – Sketch-guided Path Planning Augmented with Diffusion Experts

1. Problem Statement

Autonomous Mobile Robots (AMRs) require robust path planning that not only avoids obstacles but also adheres to nuanced human operator preferences, such as maintaining specific clearances or executing smooth turns in sensitive areas. Traditional planning algorithms (e.g., A*, RRT) struggle to encode these subjective preferences without complex reward engineering. While Imitation Learning (IL), specifically Behavioral Cloning (BC), offers a solution by learning from expert demonstrations, existing approaches face two critical limitations:

Limited Generalization: Models often fail to adapt to unseen environments or map configurations.
Data Collection Robustness: Existing annotation tools (e.g., the simulator used in the SKIPP framework) suffer from maintainability issues, lack of open-source availability, and difficulties in collecting high-quality demonstrations in specialized industrial settings.

Furthermore, while recent diffusion-based models show promise in capturing complex, multi-modal path distributions, their iterative denoising processes result in high inference latency, making them unsuitable for real-time, on-edge deployment on resource-constrained robots.

2. Methodology

The paper introduces SPADE (Sketch-guided Path Planning Augmented with Diffusion Experts), a framework designed to bridge the gap between the high representational power of diffusion models and the efficiency requirements of deployable robotic systems. The methodology consists of three primary components:

2.1 Overhauled Annotation Tool

To address data collection limitations, the authors developed a fully open-source annotation pipeline built on ROS 2.

Simulation: Utilizes map2gazebo to convert 2D occupancy grids into 3D Gazebo worlds, employing the TurtleBot3 differential-drive robot.
Navigation: Leverages the Nav2 framework for autonomous motion, allowing users to define paths interactively via RViz2 waypoints.
Data Processing: Records trajectories across path, end-goal, and robot pose modalities. A spline-based smoothing function is applied to resample paths into continuous, evenly spaced curves to reduce noise and improve training stability.

2.2 Diffusion-Augmented Behavioral Cloning (DBC)

The core innovation lies in the training strategy, which integrates a Diffusion Expert to guide a compact Behavioral Cloning (BC) network.

Architecture: The system employs a "teacher-student" paradigm. A high-capacity diffusion model (the teacher) is trained offline on expert demonstrations, while a lightweight BC model (the student) is trained for real-time inference.
Image-Conditioned Diffusion (Cond-DBC): Unlike previous DBC approaches that modeled the joint probability $p(a, s)$ $p (a, s)$ , SPADE introduces an image-conditioned diffusion model that learns the conditional probability $p(a|s)$ $p (a ∣ s)$ .
- Input: The diffusion model takes the occupancy grid, start point, and goal point as conditioning inputs (via FiLM layers) and predicts only the path channel.
- Guidance Mechanism: During BC training, the diffusion model acts as a discriminator. It calculates a Diffusion Margin Loss ( $L_{DM}$ ), defined as the difference between the reconstruction loss of the agent's predicted path and the expert's ground truth path.
- Objective: The total loss combines the standard BC loss with the diffusion margin loss ( $L_{total} = L_{BC} + \lambda L_{DM}$ ), encouraging the BC network to generate paths that align with the structural distribution of the expert diffusion model without directly mimicking its slow inference process.

2.3 Model Variants

The study evaluates three model scales (Small, Medium, Large) for the BC network, alongside two diffusion expert types:

Unconditional Diffusion (DBC): Learns joint distributions.
Conditioned Diffusion (Cond-DBC): Learns conditional distributions based on the environment state.

3. Key Contributions

Open-Source Annotation Pipeline: A robust, ROS 2-based tool that replaces deprecated simulators, ensuring reproducibility and extensibility for data collection in industrial scenarios.
Image-Conditioned Diffusion Guidance: An extension of the DBC framework that utilizes image-conditioned diffusion models to provide targeted, state-aware guidance signals, improving the interpretability of the guidance for smaller student networks.
Efficient Generalization: A training strategy that enables lightweight models to achieve "diffusion-level" generalization while retaining real-time, on-edge inference capabilities.

4. Experimental Results

The framework was evaluated on a dataset of 20,000 expert demonstrations (L-shape and U-shape paths) across various map configurations. Performance was measured using Fréchet Inception Distance (FID), Absolute Pose Error (APE), and Hausdorff Distance ( $H_{d19}$ ).

Performance vs. Baseline: The SPADE framework significantly outperformed the baseline SKIPP (standard BC) approach.
- APE: Reduced by 39.1%.
- FID: Reduced by 33.5%.
- Parameters: Achieved these gains with 93.8% fewer trainable parameters compared to large baseline models.
Cond-DBC vs. DBC: The image-conditioned approach (Cond-DBC) generally outperformed the unconditional approach (DBC), particularly in reducing artifact generation (disconnected path segments) in smaller model architectures.
Inference Efficiency: While standalone diffusion models achieved low FID scores, their inference times were prohibitive (18–45 seconds). In contrast, the medium-sized Cond-DBC model achieved comparable performance with real-time inference suitable for embedded systems.
Ablation Studies:
- Transfer Learning: Fine-tuning the SPADE pipeline from L-shape to U-shape behaviors on new maps yielded a 37.1% improvement in FID and a 62% reduction in artifacts compared to standard fine-tuning.
- Training from Scratch: The complete SPADE pipeline improved FID by 4.21% and reduced artifacts by 63.16% compared to training a standard SKIPP model from scratch.

5. Significance and Claims

The paper claims that SPADE successfully addresses the trade-off between model capacity and deployment efficiency in learning-based path planning. By using a high-capacity diffusion model as a "teacher" during training, the framework enables compact, deployable models to generalize effectively to unseen environments and behaviors without requiring cloud computation or specialized hardware accelerators.

The authors emphasize that this approach makes preference-aware path planning viable for resource-constrained industrial robots. The results suggest that diffusion-based expert guidance allows small models to capture complex, multi-modal path distributions that standard behavioral cloning cannot, while maintaining the strict latency and memory constraints required for real-world AMR deployment. The work positions SPADE as a practical step forward for integrating advanced generative priors into edge robotics without sacrificing performance.