RM2026 Engineer Assembly Algorithm

Open-source engineering assembly algorithms developed by HKUST ENTERPRIZE for the RoboMaster 2026 season. The project supports a semi-automatic energy-unit assembly workflow: the upper computer provides visual estimation, approach planning, constrained motion planning and recovery planning, while the operator performs insertion with a custom controller.

Project Contents

The repository publishes the synthetic-data pipeline, vision-model training code, numerical perception and planning algorithms, and a self-contained ROS 2 + MuJoCo integration workspace. The current runtime release targets simulation; the real-hardware backend is still being organized.

./
├── doc/                  # Open-source technical report
├── src/blender/          # Rendering and synthetic-data generation
├── src/nn/               # YOLO and LiteHRNet training/export workspace
├── runtime/              # ROS 2 Host, planning core and MuJoCo simulation
├── LICENSE
└── README.md

All relative paths and commands in the documentation assume that the current working directory is the repository root.

Environments

The runtime, neural-network training and Blender rendering workflows use separate Python environments. Their dependencies should not be installed into one shared environment:

Workflow	Environment	Setup guide
ROS 2 Host and MuJoCo runtime	ROS 2 Humble system Python at /usr/bin/python3	Runtime setup
YOLO and LiteHRNet training, inference and export	Python 3.10 virtual environment managed by uv under src/nn/.venv	Neural-network environment
Blender scene rendering	Blender 4.5.0 bundled Python 3.11	Blender environment

Blender rendering runs inside Blender's bundled Python. Dataset composition and annotation visualization run after rendering and reuse the src/nn uv environment; they do not run inside Blender. Likewise, the ROS 2 runtime uses the system Python selected by the ROS installation and must not be launched from the training virtual environment.

Documentation

• Start with the runtime setup and launch guide to run the complete ROS 2 + MuJoCo workflow.
• Technical report
• Blender data generation
• Shared neural-network environment
• YOLO pose and detection
• LiteHRNet

Dataset Release

The synthetic dataset is hosted on Hugging Face Datasets: hkustenterprize/RM26_engineer_exchange.

Split	Ratio	Samples
Train	90%	40,860
Val	10%	4,540
Total	100%	45,400

The dataset provides YOLO pose labels for the pillar and exchange targets, including 12 keypoints. A detection dataset can be derived from the same labels with the conversion command documented in the YOLO README.

Checkpoint Release

Model checkpoints are hosted on Hugging Face: hkustenterprize/RM26_engineer_exchange_model.

Model	Task	Input size	Pipeline	Native checkpoint	OpenVINO export
LiteHRNet-30-v9.1	12-keypoint	256x256	top-down	best.pth	openvino_best
YOLO26-s-v10.12	Box Detection	640x640	top-down	best.pt	openvino_best
YOLO26-s-pose-v11.02	Box Detection + 12-keypoint	640x640	bottom-up	best.pt	openvino_best

The native links are direct downloads. The OpenVINO links point to the exported model directories. All listed models were trained with synthetic data from the dataset release; the top-down pipeline was used in the deployed workflow, while the bottom-up checkpoint is provided as a reference and comparison implementation.

Release and Repository Update Log

This repository is released incrementally. The log records public artifacts and major repository reorganizations.

2026-08-13

• Consolidated the YOLO and LiteHRNet tools behind the shared rm26-nn command-line interface.
• Expanded the ROS 2 runtime, MuJoCo engine and simulation architecture documentation.
• Simplified operator input to the keyboard controls consumed by the simulation and removed an unused custom event message.
• Aligned ROS package metadata with the repository maintainer and MIT license.

2026-08-12

• Added the self-contained MuJoCo and Host ROS 2 workspace under runtime/.
• Packaged numerical perception and planning code as arm_exchange_core.
• Removed shared-memory camera transport and legacy runtime data abstractions.
• Added package-owned collision assets and simulation integration tests.

2026-08-11

• Published the Blender scene and aligned the fixed-camera rendering pipeline with Blender 4.5.0.
• Added an explicit main() entry point to render_dataset.py and a warning for exhausted framing retries.
• Consolidated the neural-network CLI and documented the shared uv environment.

2026-06-24

• Released the open-source technical report PDF.
• Added the public README and project overview.
• Added the Hugging Face dataset and checkpoint entries.

License

This repository is released under the MIT License. See LICENSE for details.

The dataset and model checkpoints are hosted on Hugging Face. Please refer to their corresponding pages for asset-specific license and usage notes.

Acknowledgement

This project was developed by HKUST ENTERPRIZE for the RoboMaster 2026 season. The implementation builds on open-source tools including MuJoCo, OMPL, hpp-fcl, Ultralytics, MMPose, SciPy and Albumentations.