# Brick Diffusion assembly policy Brick Diffusion is the visuomotor assembly policy accompanying [BrickCraft: Visuomotor Skill Composition with Situated Manual Guidance for Long-Horizon Interlocking Brick Assembly](https://arxiv.org/abs/2605.07605). It predicts end-effector pose and gripper actions to place a held target brick onto an existing structure with a single Kinova arm. See the [BrickCraft project website](https://intelligent-control-lab.github.io/BrickCraft/) for more information. ## Inputs | Field | Shape | Description | | --- | --- | --- | | `observation.images.wrist_camera` | `3 × 256 × 256` | RGB wrist-camera image | | `observation.images.fixed_camera` | `3 × 256 × 256` | Segmentation-processed fixed-camera image | | `observation.state` | `13` | FK end-effector pose (3D position + 6D rotation), gripper position, then scaled XYZ force | | `observation.task_encoding` | `6` | `[x, y, z, orientation, reference_type, target_type]` | The policy uses two consecutive observation steps. Images, robot state, and task encoding follow BrickCraft's camera cropping, segmentation, coordinate, unit, and force-scaling conventions. Input normalization uses the statistics saved with the model. ## Outputs | Field | Shape | Description | | --- | --- | --- | | `action` | `10` per step | End-effector pose (3D position + 6D rotation), then gripper position |