The paper introduces Consistency Policy, a method to accelerate visuomotor policy learning for robotic systems that are constrained by space, weight, and power limitations. Traditional visuomotor policies, such as Diffusion Policy, require high-end GPUs for fast policy inference, which is not feasible for resource-constrained robotic setups. Consistency Policy achieves faster inference by distilling a pre-trained Diffusion Policy using a consistency objective, allowing it to generate actions in a single step rather than multiple steps. This results in a significant reduction in inference time while maintaining competitive success rates. The method is evaluated on both simulated and real-world tasks, demonstrating that Consistency Policy can speed up inference by an order of magnitude compared to the fastest alternative method and outperforms it in terms of success rates. Key design choices include the choice of consistency objective, reduced initial sample variance, and the selection of preset chaining steps. The paper also discusses the robustness of the training procedure to the quality of the pre-trained Diffusion Policy and provides insights into the role of dropout in the consistency objective.The paper introduces Consistency Policy, a method to accelerate visuomotor policy learning for robotic systems that are constrained by space, weight, and power limitations. Traditional visuomotor policies, such as Diffusion Policy, require high-end GPUs for fast policy inference, which is not feasible for resource-constrained robotic setups. Consistency Policy achieves faster inference by distilling a pre-trained Diffusion Policy using a consistency objective, allowing it to generate actions in a single step rather than multiple steps. This results in a significant reduction in inference time while maintaining competitive success rates. The method is evaluated on both simulated and real-world tasks, demonstrating that Consistency Policy can speed up inference by an order of magnitude compared to the fastest alternative method and outperforms it in terms of success rates. Key design choices include the choice of consistency objective, reduced initial sample variance, and the selection of preset chaining steps. The paper also discusses the robustness of the training procedure to the quality of the pre-trained Diffusion Policy and provides insights into the role of dropout in the consistency objective.