🤖 AI

Hybrid Diffusion for Simultaneous Symbolic and Continuous Planning

This paper addresses the limitations of standard diffusion models in long-horizon robotic tasks by proposing a novel hybrid diffusion framework that simultaneously generates high-level symbolic plans and continuous trajectories, thereby improving decision-making and enabling flexible, condition-based action synthesis.

Original authors: Sigmund Hennum Høeg, Aksel Vaaler, Chaoqi Liu, Olav Egeland, Yilun Du

Published 2026-04-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sigmund Hennum Høeg, Aksel Vaaler, Chaoqi Liu, Olav Egeland, Yilun Du

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 to clean up a messy room. You want it to pick up toys, sort them by color, and put them in the right bins. This seems simple, but for a robot, it's a massive puzzle. It has to decide what to do (the strategy) and how to move its arm to do it (the physical motion) at the same time.

For a long time, scientists have used a type of AI called a Diffusion Model to teach robots. Think of a diffusion model like a "denoising" artist. If you take a clear photo and slowly add static (noise) until it's just a blur, a diffusion model learns how to reverse that process. It starts with the blur and slowly removes the noise to reveal the clear picture. In robotics, the "picture" is a smooth path for the robot's arm to follow.

The Problem: The "Blurry" Long-Term Plan
The paper explains that while these models are great at short, smooth movements (like picking up a cup), they get confused when the task is long and complex (like sorting ten blocks in a specific order).

The authors compare this to a person trying to remember a long story. If you only focus on the feelings of the story (the continuous motion) without remembering the plot points (the decisions), you might get the tone right but forget the ending. The robot ends up moving smoothly but doing the wrong things, or getting stuck in loops. It struggles to connect the "big picture" decisions with the "small picture" movements.

The Solution: The Hybrid Diffusion Planner (HDP)
To fix this, the authors created a new system called Hybrid Diffusion Planner (HDP). They realized that to solve a long puzzle, you need two things working together:

The Script (Symbolic Plan): A list of high-level steps, like "Pick up Block A," "Move to Bin 1," "Place Block A."
The Performance (Continuous Plan): The actual smooth movements of the robot's arm to execute those steps.

Instead of teaching the robot just the movements, HDP teaches it to generate both the script and the performance simultaneously.

How It Works: The "Two-Track" Denoising
The paper uses a clever trick to teach the robot both at once. Imagine you have two separate puzzles:

Puzzle A (The Motion): A blurry photo of the robot's arm moving.
Puzzle B (The Script): A sentence where some words are covered by black boxes (masked).

The HDP system learns to clean up both puzzles at the same time.

It takes the blurry motion and the covered-up script.
It uses the clues from the script to help figure out the motion. (e.g., "If the script says 'Pick up Block A,' the arm must be near Block A.")
It uses the clues from the motion to help figure out the script. (e.g., "If the arm is moving left, the next step is probably 'Move Left'.")

By letting the two puzzles talk to each other while they are being solved, the robot learns a much stronger connection between what to do and how to do it.

Why This Matters: The "Director" and the "Actor"
The authors show that this method is much better than the old ways.

Old Way: The robot tries to guess the whole path at once. It often fails, like an actor trying to improvise a whole play without a script.
New Way (HDP): The robot acts like a director and an actor working together. The "director" (the symbolic plan) gives clear instructions, and the "actor" (the motion plan) follows them precisely.

Superpowers: Following Instructions
Because the robot is generating a "script" alongside the movement, you can give it specific instructions at the last minute.

Example: You can tell the robot, "No matter what, make sure the Red Block goes on top of the tower."
The old models would ignore this or get confused.
HDP can take that instruction, lock it into the "script," and then generate the physical movements to make it happen. It's like telling a chef, "Make the pasta, but put the cheese on top," and the chef actually does it.

The Results
The team tested this in computer simulations and with a real robot arm.

In Simulations: When asked to sort an increasing number of blocks, old models failed quickly. HDP kept getting better and could handle much more complex tasks.
In the Real World: When they tried it on a real robot, HDP was much more successful at finishing the tasks without crashing or getting confused. The old models often made plans that looked okay on paper but were impossible to execute physically.

In Summary
This paper introduces a way to teach robots to plan long, complex tasks by teaching them to write a "to-do list" (symbolic plan) at the same time they learn the "dance moves" (continuous motion). By letting these two parts help each other, the robot becomes much smarter, more reliable, and easier to control, even for difficult jobs like sorting many objects or using tools.

1. Problem Statement

The paper addresses the challenge of long-horizon robotic planning using imitation learning. While Diffusion Models have become state-of-the-art for modeling continuous robotic trajectories (e.g., Diffuser), they face significant limitations when applied to complex, long-horizon tasks:

Mode Confusion: Pure diffusion models struggle to maintain coherent decision-making over long time horizons, often confusing different modes of behavior (e.g., failing to sort blocks in a specific order despite seeing successful demonstrations).
Lack of Transparency: Continuous planners operate as black boxes, making it difficult to guide or intervene in the planning process at a high level.
Inefficiency: Existing Task-and-Motion Planning (TAMP) methods are effective but require extensive hand-engineering (defining symbolic predicates, action schemas, and transition models), limiting their adaptability to new tasks via learning from demonstrations alone.

The core problem is how to combine the flexibility of diffusion-based continuous planning with the structural rigor of symbolic planning to achieve robust, interpretable, and controllable long-horizon behavior without heavy engineering overhead.

2. Methodology: Hybrid Diffusion Planner (HDP)

The authors propose the Hybrid Diffusion Planner (HDP), a framework that simultaneously generates a continuous action trajectory ( $A_c$ ) and a discrete symbolic plan ( $A_d$ ) using a coupled diffusion process.

Core Architecture

Dual Diffusion Processes: HDP employs two distinct diffusion processes running in parallel but coupled within a single neural network:
1. Continuous Diffusion (DDPM): Models the continuous robot trajectory (positions, velocities, gripper states) using standard Denoising Diffusion Probabilistic Models.
2. Discrete Diffusion (Masked Diffusion - MD4): Models the symbolic plan (e.g., "Pick Block A," "Place in Slot 1") using masked diffusion, where tokens are progressively unmasked from a fully masked state.
Joint Modeling: A unified Transformer-based denoising network ( $D_\theta$ ) takes as input the noisy continuous trajectory, the partially masked symbolic sequence, and the environment observation ( $O$ ). It predicts the noise for the continuous part and the logits for the discrete part simultaneously.
Training Objective: The model is trained on a weighted sum of losses:
- Mean Squared Error (MSE) for the continuous denoising task.
- Cross-Entropy loss for the discrete masked token prediction.
- Independent Corruption Levels: Crucially, during training, the noise levels ( $k_c$ for continuous, $k_d$ for discrete) are sampled independently. This forces the model to learn to denoise one modality using information from the other (e.g., using a clean symbolic plan to guide the denoising of a noisy trajectory, and vice versa).

Inference and Conditioning

Joint Sampling: The model generates both plans simultaneously by iteratively denoising the continuous trajectory and unmasking the symbolic tokens.
Flexible Conditioning (Inpainting): Because the symbolic plan is modeled via masked diffusion, HDP supports conditional sampling. Users can fix a partial or complete symbolic plan (e.g., "Block A must be placed first") and have the model generate a continuous trajectory that strictly adheres to these constraints. This is achieved by setting the noise level of the conditioned tokens to zero during the reverse diffusion process.

3. Key Contributions

Novel Imitation Learning Benchmark: The authors introduce a new task suite focused on long-horizon, multimodal decision-making (sorting, arranging, and tool-use) where the agent must infer the sequence of subtasks itself, rather than being given a pre-defined sequence.
Hybrid Diffusion Planner (HDP): A novel architecture that couples discrete (masked) and continuous (DDPM) diffusion processes. This allows the model to learn the correspondence between high-level symbolic logic and low-level motor control directly from data.
Flexible Controllability: The method enables fine-grained conditional planning. Unlike traditional language conditioning which requires full sentences, HDP allows users to "inpaint" specific parts of the symbolic plan, guaranteeing that the generated trajectory respects those constraints.

4. Experimental Results

The authors evaluated HDP against baselines including standard Diffuser, Hierarchical Diffuser, and variants that simply concatenate symbolic data or use separate diffusion models.

Simulation Performance:
- Success Rate: HDP significantly outperformed all baselines. On the "X-Arm Sorting" task, HDP achieved 86% success compared to 46% for the standard Diffuser.
- Scalability: As task complexity increased (e.g., sorting more blocks), baseline performance degraded monotonically. HDP remained robust, showing only a marginal drop in performance.
- Training Efficiency: HDP converged significantly faster than the standard Diffuser.
Conditional Adherence:
- When conditioned to place a specific block last, HDP adhered to the constraint 93% of the time. Baselines (Joint Diffuser, Separate Diffuser) adhered only ~32%, demonstrating that naive concatenation fails to enforce symbolic constraints.
Real-World Deployment:
- Tested on a Franka Emika robot with image-based observations. HDP maintained high success rates (70% for sorting, 60% for tool-use), whereas baselines dropped significantly (e.g., Diffuser dropped to 20% for sorting).
- Failure Modes: Baselines often failed due to "invalid plans" (logical errors in the sequence), whereas HDP failures were primarily due to "imprecise motion" (kinematic errors), indicating the high-level logic was sound.

5. Significance and Impact

Bridging the Gap: HDP successfully bridges the gap between the flexibility of learning-based continuous planners and the structural guarantees of symbolic planning, without requiring the heavy engineering of traditional TAMP systems.
Interpretability: By generating a symbolic plan alongside the trajectory, the robot's intent becomes transparent, allowing for easier debugging and human oversight.
Controllability: The ability to condition on partial symbolic plans opens new avenues for human-robot interaction, allowing operators to guide complex tasks by specifying high-level constraints rather than low-level actions.
Scalability: The method demonstrates that learning from demonstrations can scale to complex, long-horizon tasks if the model architecture explicitly accounts for the discrete decision-making structure inherent in such tasks.

In conclusion, the paper establishes that simultaneous generation of symbolic and continuous plans via hybrid diffusion is a superior approach for long-horizon robotic planning, offering robustness, interpretability, and controllability that pure continuous diffusion models lack.