Render and Diffuse (R&D) is a method that aligns image and action spaces for diffusion-based behavior cloning in robotics. It unifies low-level robot actions and RGB observations within an image space using virtual renders of the robot's 3D model. By iteratively updating these virtual renders with a learned denoising process, R&D simplifies the learning problem and introduces inductive biases that improve sample efficiency and spatial generalization. The method is evaluated in simulation and applied to six real-world tasks, demonstrating strong spatial generalization and sample efficiency compared to existing image-to-action methods. R&D aligns observation and action spaces by rendering the robot in configurations that would result from the considered actions. This allows the model to understand the spatial implications of actions and learn policies that map images to actions. The method uses a diffusion process to update the rendered actions until they closely match the training data. Several variants of R&D are introduced, each with different ways of updating actions using a denoising process. The R&D method is evaluated in simulation and real-world settings. In simulation, it is tested on 11 tasks from RLBench, showing strong performance in spatial generalization and sample efficiency. In real-world tasks, it is applied to six everyday tasks, demonstrating its effectiveness in practical scenarios. The method is also tested in a multi-task setting, where it learns multiple tasks using a single network, showing its ability to generalize across different tasks. R&D has limitations, including computational costs due to iterative rendering and forward propagation, reliance on camera calibration, and challenges with tasks involving severe occlusions and inconsistent data. However, it shows promise as a universal method for jointly representing RGB observations and actions. Future work includes extending the method to include the full robot configuration and integrating it with image foundation models.Render and Diffuse (R&D) is a method that aligns image and action spaces for diffusion-based behavior cloning in robotics. It unifies low-level robot actions and RGB observations within an image space using virtual renders of the robot's 3D model. By iteratively updating these virtual renders with a learned denoising process, R&D simplifies the learning problem and introduces inductive biases that improve sample efficiency and spatial generalization. The method is evaluated in simulation and applied to six real-world tasks, demonstrating strong spatial generalization and sample efficiency compared to existing image-to-action methods. R&D aligns observation and action spaces by rendering the robot in configurations that would result from the considered actions. This allows the model to understand the spatial implications of actions and learn policies that map images to actions. The method uses a diffusion process to update the rendered actions until they closely match the training data. Several variants of R&D are introduced, each with different ways of updating actions using a denoising process. The R&D method is evaluated in simulation and real-world settings. In simulation, it is tested on 11 tasks from RLBench, showing strong performance in spatial generalization and sample efficiency. In real-world tasks, it is applied to six everyday tasks, demonstrating its effectiveness in practical scenarios. The method is also tested in a multi-task setting, where it learns multiple tasks using a single network, showing its ability to generalize across different tasks. R&D has limitations, including computational costs due to iterative rendering and forward propagation, reliance on camera calibration, and challenges with tasks involving severe occlusions and inconsistent data. However, it shows promise as a universal method for jointly representing RGB observations and actions. Future work includes extending the method to include the full robot configuration and integrating it with image foundation models.