MTDP: Modulated Transformer Diffusion Policy Model

Qianhao Wang, Yinqian Sun, Enmeng Lu, Qian Zhang, Yi ZengarXiv 2025 · DOI · arXiv

Abstract

The Modulated Transformer Diffusion Policy (MTDP) model addresses the architectural bottlenecks in robot manipulation by redesigning how Transformers integrate environmental context. By introducing a Modulated Attention module that utilizes scale and shift parameters to fuse visual features and timesteps directly into the decoder, the authors achieved a 12% success rate increase in complex tasks. Furthermore, the integration of Denoising Diffusion Implicit Models (DDIM) nearly doubled inference speed without compromising performance.

Introduction

Behavior Cloning (BC) is a fundamental approach for robots to learn tasks from human demonstrations. While effective, it often struggles with generalization in multi-modal environments. Diffusion Policy (DP) improved upon this by using generative diffusion models to produce action trajectories. However, a new challenge emerged: standard Transformer architectures often fail to effectively integrate "guiding conditions"—the visual state of the world and the current diffusion timestep. Traditional cross-attention mechanisms frequently lack the precision required for high-stakes manipulation, leading to suboptimal performance. MTDP proposes a more "context-aware" architecture to overcome these limitations.

Methods

The Modulated Attention Framework

The core innovation of MTDP is the Modulated Attention module. Inspired by the Diffusion Transformer (DiT), this approach moves beyond simple cross-attention to inject information more deeply into the network's layers.

  1. Encoder-Decoder Architecture: The model employs an encoder to process image features and timesteps, and a decoder to generate the final action trajectories.
  2. Modulation Parameters: The encoder's output is fed through a Multi-Layer Perceptron (MLP) to generate specific scale and shift parameters ($\gamma, \beta$).
  3. Layer-Wise Integration: These parameters modulate the LayerNorm and attention outputs within the decoder. This ensures that every part of the decoder is continuously informed by the environmental context.
Overview of the MTDP training and generation pipelines, illustrating the flow from image features and noised actions to predicted noise and final action trajectories.

Architectural Variants

The authors explored several configurations to optimize performance. Through ablation studies, they identified the M-SelfAttention variant as the most effective. In this configuration, the self-attention layers of the decoder—where the action trajectory "communicates" with itself—are modulated, while standard cross-attention is maintained for encoder features. This provides a tighter coupling between the robot's goal and its generated actions.

Detailed architectural diagrams of the proposed modulation modules, including the high-performing M-SelfAttention variant.

Inference Optimization (MTDP-I)

To make the model viable for real-time control, the authors swapped the standard Denoising Diffusion Probabilistic Model (DDPM) for a Denoising Diffusion Implicit Model (DDIM). This allowed for a reduction in sampling steps from 100 to 60, nearly doubling generation speed while maintaining success rates higher than non-modulated baselines.

Key Results and Insights

MTDP was evaluated on six standard robot manipulation benchmarks, including PushT, Toolhang, and Transport.

  • Superior Success Rates: MTDP outperformed the baseline DP-Transformer across nearly all tasks. Most notably, in the highly complex Toolhang task (assembling a tool rack), MTDP achieved a 12% improvement in success rate.
  • Universal Applicability: The modulation logic was also successfully applied to a UNet architecture (MUDP), demonstrating that better context integration is a general benefit for diffusion-based robotics.
  • Efficiency: The MTDP-I variant demonstrated that architecture-level optimizations allow for significantly fewer iterations during inference, a critical factor for real-world robotic deployment.

Conclusion

MTDP demonstrates that the method of context integration is just as important as the model's capacity. By replacing standard cross-attention with a modulated approach, the authors have created a more robust and efficient policy for complex robot manipulation. Future research will focus on scaling this mechanism to Vision-Language-Action (VLA) models, enabling robots to follow natural language instructions with the same high-precision control.

Summary updated

BibTeX
@misc{wang2025mtdp,
  author = {Qianhao Wang and Yinqian Sun and Enmeng Lu and Qian Zhang and Yi Zeng},
  title = {{MTDP: Modulated Transformer Diffusion Policy Model}},
  year = {2025},
  eprint = {2502.09029},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.09029},
}
Figure