BrickCraft model checkpoints

Brick Diffusion assembly policy and task-specific YOLOv8n-seg checkpoints for the single-arm BrickCraft assembly pipeline.

• Project website
• Paper

Included checkpoints

Path	Purpose	Class IDs
brickcraft_diffusion/pretrained_model/	Predict end-effector pose and gripper actions for single-step assembly and the assembly skill in long-horizon execution	Not applicable
yolo/seg_structure.pt	Segment the assembled brick structure for scene alignment	0: object
yolo/seg_brick_gripper.pt	Segment the target brick and gripper	0: brick, 1: gripper

See the YOLO checkpoint notes for loading instructions and provenance. The checkpoint metadata records Ultralytics version 8.4.8; packaging did not include a compatibility test with other versions.

See the Brick Diffusion checkpoint notes for the policy's functionality, inputs, and outputs. Keep its saved preprocessing and postprocessing files with the weights.

Download

hf download JerryYuu/BrickCraft --local-dir ./models/BrickCraft

From the BrickCraft code repository root, configure:

{
  "yolo_seg_structure_ckpt": "./models/BrickCraft/yolo/seg_structure.pt",
  "yolo_seg_brick_gripper_ckpt": "./models/BrickCraft/yolo/seg_brick_gripper.pt"
}

For reproducibility, pin --revision to the model repository commit used in your experiment.

For policy inference, set the assembly controller's pretrained_path to ./models/BrickCraft/brickcraft_diffusion/pretrained_model. See the project website for usage resources.

Scope and verification

This repository contains the two YOLO checkpoints and the Brick Diffusion policy. SAM-2 is not included. These files alone do not provide the complete robot execution stack. The policy supplies the assembly skill; the long-horizon pipeline uses a separate scripted pick skill. Skill failures still require a human decision.

Policy weights, model configuration, processor files, and YOLO checkpoints are preserved without modification. The policy training configuration uses release names for its dataset path, output directory, and job name; other fields are unchanged. Unmodified policy files were compared with their source using SHA256, and their JSON configuration, processor state references, and safetensors headers were checked. The YOLO files retain their earlier ZIP archive integrity checks. SHA256SUMS covers the included files. Inference accuracy, CUDA execution, and physical robot operation were not tested during this packaging step.

License

The YOLO checkpoints retain their AGPL-3.0 notice; see LICENSE and the YOLO checkpoint notes. The Brick Diffusion checkpoint's saved configuration does not specify a model license (license: null).