Factored Diffusion Policies:Compositionally Generalized Robot Control with a Single Score Network

2026-05-21 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Factored Diffusion Policies introduce a novel approach to robot control, utilizing a single shared diffusion network trained with per-factor null-token dropout. This method allows the network's score to decompose additively across factors during inference, significantly reducing the training-task budget from a combinatorial product to a sum of factor cardinalities. The approach relies on approximate conditional independence between factors given the action-observation pair, providing a bounded uniform error. A trajectory-tube certificate further chains this score-level bound through the reverse-time sampling ODE and a contracting tracking controller. Drone racing experiments confirm the generalization bound and certificate, with the factored policy passing 90% of held-out gates, matching an oracle, while a K-network composition baseline achieved only 3%. It also demonstrated zero-shot transfer to unseen venues, yielding an +11.7pp success-rate gain and a 2.4X crash-rate reduction.

Key takeaway

For Robotics Engineers developing compositionally generalized robot control systems, you should consider Factored Diffusion Policies to significantly reduce training data requirements and improve zero-shot generalization. This approach offers superior performance over multi-network baselines, achieving 90% success on held-out tasks and 2.4X crash reduction in vision-based scenarios. Evaluate its applicability for tasks with multiple, independently varying factors to streamline development and deployment.

Key insights

Factored Diffusion Policies use a single diffusion network with additive score decomposition for compositionally generalized robot control.

Principles

• Additive score decomposition generalizes robot control.
• Per-factor null-token dropout enables compositional learning.
• Approximate conditional independence bounds error.

Method

Train a single diffusion network using per-factor null-token dropout. At inference, decompose the score additively across factors to approximate the true joint score.

In practice

• Apply to multi-gate drone racing.
• Achieve zero-shot transfer to new environments.
• Reduce crash rates in vision-based tasks.

Topics

• Factored Diffusion Policies
• Robot Control
• Compositional Generalization
• Diffusion Networks
• Drone Racing
• Zero-shot Transfer

Best for: Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• Ambient Diffusion Policy: Imitation Learning from Suboptimal Data in Robotics — Ambient Diffusion Policy leverages noise-dependent data usage to extract useful features from suboptimal robotic demonstration data.
• Physical Intelligence shows robot model with LLM-like generalization, flaws included — π0.7 is a robot foundation model exhibiting compositional generalization by recombining learned skills.
• Learning Altruistic Collaboration in Heterogeneous Multi-Team Systems — A GNN policy extends Hamilton's rule for altruistic robot allocation in heterogeneous multi-team systems, solving NP-hard problems scalably.
• Proposal-Conditioned Latent Diffusion for Closed-Loop Traffic Scenario Generation — A new diffusion framework improves closed-loop traffic simulation efficiency and control for autonomous vehicles.
• From Seeing to Simulating: Generative High-Fidelity Simulation with Digital Cousins for Generalizable Robot Learning and Evaluation — WorldComposer generates diverse, high-fidelity simulation environments from real-world panoramas for robust robot learning.
• Diverse Yet Consistent: Context-Guided Diffusion with Energy-Based Joint Refinement for Multi-Agent Motion Prediction — Context-guided diffusion with energy-based refinement improves diverse yet consistent multi-agent motion prediction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.