Towards Human-Centered Automation: An RL-Driven Robotic Framework for Humanely Supporting Labor-Intensive Construction Workers

In preparation, doctoral dissertation.

Yuning Wu

Committee: Daniel Cardoso Llach (Advisor), Jean Oh (Co-Advisor), Jieliang Luo
I am currently in Ph.D. ABD (All But Dissertation) stage, during which I am mostly focusing on final dissertation writing.

Towards Human-Centered Construction Robotics: An RL-Driven Companion Robot For Contextually Assisting Carpentry Workers

Under review, submitted to IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).

Yuning Wu, Jiaying Wei, Jean Oh, Daniel Cardoso Llach

/ PDF / arXiv / video

TL;DR: This paper documents the technical development of a socially-aware 'work companion robot' centered on assisting, rather than brutally replacing, human workers. By taking on the more physically burdensome tasks within an existing prototypical construction scenario, carpentry formwork, the robot aims to allow workers to focus on the more high-skilled and sophisticated facets of their daily practice. The paper offers a systematic overview of how the robot prototype is architected, driven by state-of-the-art (SoTA) reinforcement learning algorithms. It also proposes a novel pipeline for incrementally fine-tuning a pretrained RL model to align with context-specific worker behaviors and prevent drastic policy shifts during Sim2Real.

Construction as Computer-Supported Cooperative Work: Designing a Robot for Assisting Carpentry Workers

In revision, submitted to ACM Conference on Computer Supported Cooperative Work (CSCW).

Yuning Wu*, Emek Erdolu*, Jiaying Wei, Jean Oh, Daniel Cardoso Llach

TL;DR: This paper reimagines the integration of robotics into highly manual construction work from a socio-technical perspective. It emphasizes the application of Computer-Supported Cooperative Work (CSCW) principles, presenting a case study on the design, development, and evaluation of a prototype robot intended to assist in carpentry formwork. This paper discusses the use of ethnographic and practice-based research methods to understand the cooperative work of construction teams and inform the design of supportive robot technologies. It proposes specific human-centered benchmarking metrics for evaluating the envisioned robot support and advocates for an approach that supports rather than completely reconfigures or replaces the human elements in existing labor-intensive work.

Correcting Sparse Erroneous Navigation Behaviors through Few-Shot Learning from Two Teachers

In preparation.

Yuning Wu, Jean Oh

TL;DR: This research introduces a novel approach to address long-tailed errors in robot navigation, utilizing a Vision Language Model (VLM) and human expert preferences for corrections. By leveraging CLIP as few-shot learners, the method effectively compensates for the sparse data issue inherent in long-tailed distributions. The paper proposes a new architecture aimed at efficiently rectifying erroneous navigation behaviors and better aligning them with human preferences, offering a significant advancement in social navigation systems. This paper is currently in preparation. It will be submitted to a top-tier conference in robotics and AI.

Learning Dense Reward with Temporal Variant Self-Supervision

[ICRA'22] IEEE International Conference on Robotics and Automation 2022 | RL for Contact-Rich Manipulation Workshop.

Yuning Wu, Jieliang Luo, Hui Li

/ PDF / arXiv / slides / code

TL;DR: This paper introduces an efficient framework for deriving dense rewards in reinforcement learning, focusing on enhancing the efficiency and robustness of training for robots performing contact-rich manipulation tasks. Developed through collaboration between Carnegie Mellon University and Autodesk Research, the framework leverages Temporal Variant Forward Sampling (TVFS) and a self-supervised learning architecture. By utilizing temporal variance and multimodal observation pairs, the generated dense reward can facilitate faster and more stable policy training in a plug-and-play manner, examplified in tasks such as joint assembly and door opening. This work aims to share new thinkings in reward learning and its applications in robotics.

Towards a Distributed, Robotically-Assisted Construction Framework: Using Reinforcement Learning to Support Scalable Multi-Drone Construction in Dynamic Environments

[ACADIA'20] Annual Conference of the Association of Computer Aided Design in Architecture 2020.

Zhihao Fang, Yuning Wu, Ammar Hassonjee, Ardavan Bidgoli, Daniel Cardoso Llach

/ PDF / CumInCAD / slides / video / code

TL;DR: The paper explores an innovative framework that utilizes multi-agent reinforcement learning method (MADDPG) to enable drones to perform collaborative tasks such as bricklaying and spray-coating. This scalable, adaptive approach allows multiple drones to work in dynamic environments, promising to enhance efficiency and flexibility in construction processes. By demonstrating the framework's potential through simulations and advancing towards a hardware prototype, the study contributes to the evolving field of construction robotics by offering insights into the future of human-machine collaborative ecosystems.

Teaching an AI Agent Spatial Sense by Playing LEGO

Under submission.

Yuning Wu, Katerina Fragkiadaki

TL;DR: Whether an architect or not, how we build reflects our spatial sense of the physical world. As children, this exploration typically begins with playing with toys like building blocks. This research traces this initial baby-step learning process by training an AI agent to play with LEGOs in a physical simulator (NVIDIA Omniverse Isaac Gym / PyBullet). A key aspect is enabling the AI agent to gain a spatial sense by roughly understanding whether a composite structure is stable, through grasping the relationships between blocks. This paper is currently under submission to a top-tier conference in AI.

State Space Paradox of Computational Research in Creativity

[HI'19] Thirtieth Anniversary 'Heron Island' Conference on Computational and Cognitive Models of Creative Design.

Ömer Akin, Yuning Wu

/ PDF / arXiv / slides

In memory of Professor Ömer Akin (1942-2020). This is one of his last papers. I am priviledged to be organizing and presenting it. This paper explores the State Space Paradox (SSP) in computational creativity research, discussing how current digital systems, designed as closed systems with predefined parameters, are inherently limited in their creative capacity compared to the open systems of human creativity. It examines various procedural and representational approaches to artificial creativity, such as genetic algorithms and shape grammars, and critiques their ability to truly emulate human creativity. The paradox lies in the fact that while these systems can produce outputs that may initially seem creative, they are unable to exceed their predefined state spaces and thus fail to achieve genuine novelty or adapt beyond their initial programming.