Hugging Face Proposes EAPN: Enhancing Consistency in Asynchronous Replanning for Generative Robot Policies

Summary:Hugging Face's research team introduces Execution-Aligned Progressive Noise (EAPN), a novel method to address mode switching and inconsistency issues in asynchronous replanning for generative robot policies. EAPN propagates a shared noise trajectory across replanning steps and aligns it with actual execution displacement, establishing execution-aligned inter-chunk correlation. Within each action chunk, it models temporal correlation along action time and combines the aligned stochastic history w

---

Background and Challenges

Continuous asynchronous replanning is essential for real-time generative robot policies. However, independent stochastic initialization can lead to mode switching and inconsistency across action chunks, affecting the coherence and reliability of robot task execution.

Overview of EAPN

Hugging Face's research team introduces Execution-Aligned Progressive Noise (EAPN), a novel method that addresses these issues by introducing structured stochasticity. The key innovations of EAPN include:

• Inter-Chunk Correlation: Propagates a shared noise trajectory across replanning steps and aligns it with actual execution displacement, establishing execution-aligned inter-chunk correlation.
• Temporal Correlation Modeling: Models temporal correlation along action time within each action chunk to ensure coherence within the chunk.
• Execution-Consistent Generative States: Combines the aligned stochastic history with committed action context to condition subsequent generation, allowing new chunks to continue from execution-consistent states rather than restarting from independent noise.

Experiments and Results

EAPN has been evaluated across multiple benchmarks, including D3IL, Kinetix, and LIBERO, as well as real-world manipulation tasks:

• D3IL: EAPN significantly improves multimodal behavior consistency.
• Kinetix: Achieves an average success rate of 88.59%.
• LIBERO: Remains robust even under long inference delays, maintaining strong task performance.
• Real-Robot Experiments: Achieves 90.0% success in object storage tasks and 96.7% in bimanual cloth folding tasks.

Technical Highlights

1. Structured Stochasticity Introduction: Ensures the coherence and consistency of the generation process through the introduction of stochasticity at both inter-chunk and intra-chunk levels.
2. Execution Alignment Mechanism: Solves inconsistency issues in asynchronous replanning by aligning the noise trajectory with actual execution displacement.
3. Efficient Training and Inference Workflow: EAPN maintains high success rates without significantly increasing computational overhead, making it suitable for practical deployment scenarios.

Industry Impact and Developer Recommendations

EAPN provides a new technical path for asynchronous replanning in robot policies, particularly excelling in multimodal behavior consistency and task performance under long inference delays. Developers can refer to EAPN's design principles for applications and innovations in areas such as robot control and intelligent agent decision-making. Additionally, EAPN's release offers new research directions for modeling complex dynamic systems and real-time control tasks.

---

Source: Hugging Face Daily Papers (2026-10-05)

— END —

Tags: #Hugging Face Follow #Robot Policy Follow #Asynchronous Replanning Follow #EAPN Follow #Multimodal Behavior Follow