Abstract
This paper presents the first comprehensive application of legal-action masked Deep Q-Networks with safe ZYZ regrasp strategies to an underactuated gripper-equipped 6-DOF collaborative robot for autonomous Soma-cube assembly learning. We address three critical challenges in robotic manipulation: combinatorial explosion in action spaces, unsafe motion planning, and systematic assembly-strategy learning. Our system integrates a legal-action-masked DQN with hierarchical architecture that decomposes Q-function estimation into orientation and position components, reducing computational complexity from O(3, 132) to O(116) + O(27) while maintaining solution completeness. Curriculum learning across three progressive difficulty levels (2-piece, 3-piece, 7-piece) achieves training efficiency of 100% for Level 1 within 500 episodes, 92.9% for Level 2, and 39.9% for Level 3 over 105,300 total training episodes. ZYZ singularity guards prevent gimbal lock, improving motion success from 54% to 96%. Real-time perception via Unity-based global mapping processes 300,000 points at 30 FPS with Intel RealSense D435i. Human-robot collaboration through Whisper-based speech recognition achieves 94% accuracy for Korean commands. Extensive experimental validation on a Doosan M0609 demonstrates a production-ready platform advancing intelligent collaborative robotics.
A full production-grade collaborative-robotics system built around
the Soma-cube assembly task, demonstrating that disciplined action
masking, singularity-safe regrasp planning, and multimodal HRI can
be composed into a deployable platform. First published on arXiv as
2508.21272.
- Legal-action masking reduces the action space from 4,536 → 2,484
feasible actions — a 26% sample-efficiency improvement with no loss
of solution completeness.
- ZYZ regrasp with proximity-based singularity detection prevents
gimbal lock, raising motion success 54% → 96%.
- Sim-to-real bridge — 75% assembly success rate with ±1.8 mm
positioning accuracy in manufacturing-relevant conditions.
- Curriculum learning achieves 100% / 92.9% / 39.9% success across
2-piece, 3-piece, and 7-piece levels.
- Korean-language HRI — Whisper-based speech recognition at 94%
accuracy.
Jaehong Oh, Seungjun Jung, Sawoong Kim — Doosan Robotics Rokey
Bootcamp, Seoul. Work supported by K-Digital Training Program,
mentored by Chunghyeon Lee.
The most hands-on paper in the collection. It exercises the
cognitive-robotics stack end-to-end on real hardware and links
tightly to the SEGO architecture.
@misc{oh2025soma,
title = {Learning to Assemble the Soma Cube with Legal-Action Masked DQN and Safe ZYZ Regrasp on a Doosan M0609},
author = {Jaehong Oh and Seungjun Jung and Sawoong Kim},
year = {2025},
eprint = {https://arxiv.org/abs/2508.21272},
archivePrefix = {arXiv},
url = {https://jack0682.github.io/papers/soma-cube-assembly-dqn/},
}