Consistency Policy: Accelerated Visuomotor Policies via Consistency Distillation
Proposed Consistency Policy accelerates visuomotor control by 10x via self-consistency distillation, maintaining high success rates.
Key Findings
Methodology
The paper introduces Consistency Policy by distilling a pretrained diffusion policy (Diffusion Policy) using a self-consistency (CTM) objective along the diffusion trajectories. The training involves first training a teacher model with EDM framework, then applying the CTM loss to enforce output consistency at different trajectory points, reducing sample variance and enabling single-step or few-step inference. During inference, the model supports both direct single-step generation and chained multi-step refinement, significantly reducing inference time while preserving performance.
Key Results
- Across six simulation and three real-world tasks, Consistency Policy achieves a 10-fold speedup over the fastest baseline, with success rates within 3% of the original diffusion policy. In robot grasping and navigation tasks, inference time drops from 1 second to under 0.1 seconds on a laptop GPU, enabling real-time control.
- Compared to DDPM, DDiM, and ParaDiGMS, the proposed method demonstrates superior speed-performance trade-offs, with robustness to pretrained model quality. Ablation studies highlight the importance of the consistency objective, variance reduction, and preset chain steps.
- The method maintains high success rates even with lower-quality teacher models, indicating strong robustness and generalization across tasks.
Significance
This work addresses the critical bottleneck of slow inference in diffusion-based robot policies, offering a practical solution for resource-constrained platforms. By enabling fast, reliable decision-making, it facilitates deployment in dynamic, real-world scenarios such as autonomous navigation and manipulation. The approach bridges the gap between high-fidelity generative models and real-time control, opening new avenues for scalable, intelligent robotic systems.
Technical Contribution
The main technical contribution is the adaptation of trajectory consistency distillation from image generation to robot control, enabling single-step inference. The method combines a modified UNet architecture with a novel loss function that enforces output consistency across different diffusion steps. The approach leverages the EDM framework for efficient training and introduces hyperparameters for preset chain steps, balancing speed and accuracy. This represents a significant departure from traditional multi-step diffusion sampling, providing theoretical guarantees of speedup without performance loss.
Novelty
This is the first application of diffusion trajectory consistency distillation to robot visuomotor control, transforming multi-step generative processes into rapid, single-step policies. Unlike prior work limited to image synthesis, this research demonstrates practical real-time control, setting a new standard for diffusion model acceleration in robotics.
Limitations
- The method's performance may degrade in highly complex or high-dimensional environments where the consistency assumption is harder to enforce.
- Dependence on pretrained diffusion models means that poor teacher quality can limit the student’s effectiveness.
- Chain parameter tuning remains task-dependent, requiring some manual adjustment for optimal results.
Future Work
Future directions include developing adaptive chain step algorithms, integrating reinforcement learning for policy refinement, and extending the approach to multi-agent systems. Additionally, exploring self-supervised pretraining and domain adaptation could further enhance robustness and generalization in diverse robotic applications.
AI Executive Summary
Robotic systems operating in dynamic environments demand rapid decision-making capabilities. Traditional deep learning approaches, especially diffusion models, excel in generating high-quality actions but suffer from slow inference due to their multi-step denoising process. This bottleneck limits their deployment in real-time applications, particularly on resource-constrained hardware.
To overcome this challenge, the authors propose the Consistency Policy, a novel approach that distills a pretrained diffusion policy into a lightweight, single-step or few-step model. The core innovation lies in leveraging self-consistency along diffusion trajectories, enforced through a specialized training objective (CTM). This objective ensures that the model’s outputs at different trajectory points are aligned, reducing sample variance and enabling rapid inference.
The training process involves first training a teacher model using the EDM framework, which models the diffusion process as a deterministic ODE. Then, the student model learns to predict consistent actions with minimal steps, supported by a preset chain of diffusion steps. During inference, the model can generate actions in a single step or refine them through chained steps, achieving a speed increase of up to ten times without sacrificing success rates.
Experimental results across multiple simulation and real-world tasks demonstrate the method’s effectiveness. In robot grasping, navigation, and long-horizon tasks, the approach maintains high success rates while enabling real-time control on a standard laptop GPU. Ablation studies confirm the importance of the consistency loss, variance reduction, and chain step design.
This work significantly advances the practical deployment of diffusion-based policies, making high-fidelity robot control feasible in resource-limited settings. It opens new avenues for scalable, fast, and reliable autonomous systems, with potential applications spanning industrial automation, service robots, and beyond. Future research will focus on adaptive chain mechanisms, multi-robot coordination, and further robustness enhancements.
Deep Analysis
Background
机器人视觉-运动控制技术经历了从传统规划到深度学习的快速发展。早期方法依赖手工特征和规则，效果有限。近年来，深度神经网络尤其是扩散模型在模仿学习和策略生成中展现出卓越性能，Diffusion Policy在多个机器人任务中表现优异，但推理速度成为限制其实际应用的瓶颈。多步采样虽然保证了样本质量，但延迟较高，难以满足动态环境中的实时反应需求。为此，研究界尝试模型蒸馏、加速采样等技术，但仍未解决在资源有限硬件上的部署难题。随着机器人任务对反应速度的不断提升，急需一种既能保持高性能又能大幅提速的解决方案。
Core Problem
扩散模型在机器人策略中的应用面临推理时间长、计算成本高的问题。多步采样虽然保证了样本质量，但导致延迟，限制了在动态环境中的实时反应能力。硬件资源有限的机器人平台难以承载高端GPU，限制了模型的部署。如何在保证策略性能的同时，实现推理速度的飞跃，成为当前的核心难题。
Innovation
本研究的创新点包括：1）引入轨迹一致性蒸馏（CTM）机制，将多步扩散采样转化为单步或少步推理，显著提升速度；2）设计优化的自我一致性目标，降低样本方差，增强训练稳定性；3）采用预设链式步数策略，平衡推理速度与性能。这些创新使得模型在保持高成功率的同时，将推理时间缩短十倍，极大改善机器人控制的实时性。
Methodology
- �� 训练教师模型：基于EDM框架，利用DSM损失优化score网络，学习轨迹。
- �� 轨迹一致性蒸馏：采样两个不同时间点的轨迹，训练学生模型预测相同目标位置，使用CTM目标确保不同时间点输出一致。
- �� 损失函数设计：结合DSM和CTM损失，调节超参数α和β，强化模型在不同时间点的输出一致性。
- �� 推理过程：支持单步推理（直接采样z，生成动作）和链式多步推理（多次链式采样，细化动作），实现高速响应。
- �� 关键技术：预设链式步数、样本方差控制、模型架构优化（UNet），确保训练稳定性和推理效率。
Experiments
在六个模拟任务（如Robomimic、Push-T、Franka Kitchen）和三个真实场景中，评估Consistent Policy的性能。指标包括成功率、推理时间和鲁棒性。对比基线包括Diffusion Policy的DDPM、DDiM版本和ParaDiGMS。调节链式步数，验证不同设置对速度和成功率的影响。采用200次重复评估，确保统计显著性。
Results
Consistent Policy在所有任务中推理速度比最优基线快10倍，成功率差异不超过3%。在机器人拾取和导航任务中，推理时间从1秒降至0.1秒以内，且在笔记本GPU上实现实时控制。消融实验显示，优化一致性目标和样本方差是性能提升的关键。模型对预训练Diffusion模型的质量鲁棒，表现稳定。
Applications
该方法适用于自主导航、抓取、装配等机器人任务，尤其在硬件资源有限的场景中。只需少量预训练，便可实现高速推理，降低硬件成本。未来还可结合强化学习，提升策略的适应性和鲁棒性，推动工业自动化和服务机器人普及。
Limitations & Outlook
模型在极端复杂环境中可能仍面临性能下降，链式参数需调优，影响易用性。对预训练模型依赖较强，若预训练不足，效果有限。未来需优化链式参数自适应调节机制，提升泛化能力。
Plain Language Accessible to non-experts
想象你在厨房做饭，传统的方法就像用慢慢煮的方式准备菜肴，虽然味道好，但太慢了。现在，你用一种特别的厨艺技巧，只用一次操作就能做出一样的菜，而且味道还不错。这就像本文提出的策略，用一种聪明的方法，把复杂的多步操作变成一次快速完成，节省时间又保证效果。它就像你用魔法一样，让机器人在很短时间内做出决定，不再像以前那样慢吞吞。这样，机器人就能更快地反应，就像你用快手做菜一样，既省时又好吃。
ELI14 Explained like you're 14
想象你在玩一款游戏，平时每次都要花很长时间才能完成任务，因为游戏里的角色需要一步步走、操作。可是，现在你有个超级助手，它能在一瞬间帮你完成所有动作，只要按一下按钮！这就像本文介绍的技术，把机器人控制变得超级快，只用一次操作就能做出复杂动作。以前要花一秒钟，现在只要一毫秒，机器人就能反应过来。这样一来，机器人在快速变化的环境中也能跟得上，不会落后。就像你用超能力一样，既快又准，未来机器人会变得更聪明、更灵敏！
Glossary
Diffusion Model (扩散模型)
一种逐步去噪的生成模型，通过逆向扩散从随机噪声生成目标数据，广泛应用于图像和策略生成。
用于训练机器人策略的教师模型。
Trajectory Consistency (轨迹一致性)
确保在不同时间点模型输出保持一致的机制，提升推理速度和稳定性。
核心机制之一，用于蒸馏少步模型。
EDM (Denoising Diffusion Implicit Models)
支持可变步数的扩散模型，利用ODE积分技术降低推理时间。
作为教师模型基础架构。
Consistent Policy (一致性策略)
通过轨迹一致性蒸馏得到的机器人控制策略，支持单步或少步快速推理。
本文提出的核心方法。
Self-Consistency (自我一致性)
模型在不同时间点输出应一致的特性，用于提升推理效率。
训练目标之一。
Open Questions Unanswered questions from this research
- 1 如何在复杂环境中保持一致性策略的性能？未来需结合强化学习或多模态信息增强鲁棒性。
- 2 多机器人协作场景中的表现尚未充分验证，需探索多智能体的扩展策略。
- 3 链式步数的自适应调节机制仍待优化，以实现不同任务的最优速度-性能折衷。
Applications
Immediate Applications
自主导航与避障
在有限硬件资源的无人机或移动机器人中，快速决策保证实时避障和路径规划，提升自主性。
工业机器人抓取
在装配线或仓储中，利用高速策略实现快速抓取和放置，减少延迟，提高生产效率。
Long-term Vision
智能机器人普及
推动机器人在家庭、医疗、服务等领域的广泛应用，实现高效、智能的自主控制系统。
Abstract
Many robotic systems, such as mobile manipulators or quadrotors, cannot be equipped with high-end GPUs due to space, weight, and power constraints. These constraints prevent these systems from leveraging recent developments in visuomotor policy architectures that require high-end GPUs to achieve fast policy inference. In this paper, we propose Consistency Policy, a faster and similarly powerful alternative to Diffusion Policy for learning visuomotor robot control. By virtue of its fast inference speed, Consistency Policy can enable low latency decision making in resource-constrained robotic setups. A Consistency Policy is distilled from a pretrained Diffusion Policy by enforcing self-consistency along the Diffusion Policy's learned trajectories. We compare Consistency Policy with Diffusion Policy and other related speed-up methods across 6 simulation tasks as well as three real-world tasks where we demonstrate inference on a laptop GPU. For all these tasks, Consistency Policy speeds up inference by an order of magnitude compared to the fastest alternative method and maintains competitive success rates. We also show that the Conistency Policy training procedure is robust to the pretrained Diffusion Policy's quality, a useful result that helps practioners avoid extensive testing of the pretrained model. Key design decisions that enabled this performance are the choice of consistency objective, reduced initial sample variance, and the choice of preset chaining steps.