Tao Sun

Open source

Code · 2026

AssembleNIST: Assembly Policy Learning in MuJoCo

Tao Sun

Open-source code release · MIT License

Code Videos BibTeX
A simulated Franka Panda arm assembling parts on the NIST Assembly Task Board 1 in MuJoCo.
One policy assembling parts on the NIST Assembly Task Board 1. The inset shows the wrist camera.

AssembleNIST is an independent, open-source implementation of AlphaNIST in MuJoCo. A single reinforcement-learning policy, trained with PPO from binary success rewards, controls a Franka Panda arm and completes all 20 assembly tasks on the NIST Assembly Task Board 1. Code, pretrained policies, and instructions are in the GitHub repository. This project is not affiliated with or endorsed by the AlphaNIST authors.

Results

With 8,192 parallel environments, the policy reaches 98.4% success after 10K PPO iterations and 98.7% after 12.5K. AlphaNIST reports 86.9% after 25K iterations, but it uses a different simulator, asset set, and implementation, so the numbers are not directly comparable.

Success rate over training iterations. It rises from about 25% to above 98% by 5,500 iterations. AlphaNIST’s reported results are marked at 25K iterations.

The 20 parts at random initial poses on the board. The board after the policy assembles every part.

Left: the 20 parts at random initial poses. Right: the board after assembly.

Acknowledgements

This project builds on AlphaNIST and Success Guided Sampling. Thanks to Octi Zhang for guidance on training details and Jing Zhang for helpful discussions.

BibTeX

@misc{sun2026assemblenist,
  title        = {AssembleNIST: One Policy for All NIST Board Assemblies in MuJoCo},
  author       = {Sun, Tao},
  year         = {2026},
  howpublished = {\url{https://github.com/suniique/assembly-nist}}
}