[Paper Review] Coevo: a collaborative design platform with artificial agents

Gerard Serra, David Miralles|arXiv (Cornell University)|Apr 30, 2019

Social Robot Interaction and HRI17 references4 citations

TL;DR

Coevo is an online collaborative design platform enabling humans and artificial agents to co-create 2D shapes using shared tools and a common design language. By combining evolutionary computation with interactive, transparent design processes in physics-simulated environments, Coevo enables real-time knowledge transfer, explainable AI behavior, and the generation of diverse, high-performing design solutions.

ABSTRACT

We present Coevo, an online platform that allows both humans and artificial agents to design shapes that solve different tasks. Our goal is to explore common shared design tools that can be used by humans and artificial agents in a context of creation. This approach can provide a better knowledge transfer and interaction with artificial agents since a common language of design is defined. In this paper, we outline the main components of this platform and discuss the definition of a human-centered language to enhance human-AI collaboration in co-creation scenarios.

Motivation & Objective

• To develop a shared design environment where humans and artificial agents use identical tools and a common language to co-create solutions.
• To enhance knowledge transfer and mutual understanding in human-AI collaboration by making AI design processes transparent and modifiable.
• To explore how population-based optimization algorithms can generate novel, high-performing design solutions in shared, interactive environments.
• To support incremental learning and iterative design by preserving and visualizing the entire design process for both humans and agents.
• To expand the solution space in design tasks by enabling diverse, unexpected, yet valid proposals through evolutionary computation.

Proposed method

• Designs are constructed from 2D bricks using basic actions: add, remove, rotate, and connect, forming continuous 2D shapes.
• The platform uses p5.js and matter.js to render interactive, physics-based 2D environments for real-time simulation and evaluation.
• Four distinct design challenges (collect, move, cut, protect) define tasks with quantifiable performance scores (0–1) based on objective criteria.
• Artificial agents use a population-based optimization algorithm that evolves designs based on performance feedback from the environment.
• All design actions—by humans or agents—are logged and visualized, enabling inspection, editing, and continuation at any stage.
• The system supports interactive intervention: users can modify any component during optimization, maintaining continuity and transparency.

Experimental results

Research questions

• RQ1 How can a shared design language and common tools enable effective collaboration between humans and artificial agents in creative tasks?
• RQ2 To what extent can population-based evolutionary algorithms generate diverse, high-performing, and novel design solutions in interactive environments?
• RQ3 How does real-time, transparent design process logging improve human understanding and trust in AI-generated designs?
• RQ4 Can interactive, incremental design with human-AI co-evolution lead to broader exploration of the solution space compared to black-box AI systems?
• RQ5 What role does shared evaluation criteria play in enabling fair comparison and knowledge transfer between human and AI design processes?

Key findings

• The Coevo platform successfully enables both humans and artificial agents to use identical tools and design actions, creating a shared design language.
• Artificial agents generated multiple valid, diverse solutions for the same task—such as moving on an inclined plane—demonstrating novel, non-intuitive designs beyond the common wheel.
• The system's transparent logging of all design steps allows full inspection and mid-process editing by users, enhancing explainability and user control.
• Performance scores (0–1) based on objective criteria enabled consistent evaluation and iterative improvement for both human and AI-generated designs.
• Population-based optimization led to broader exploration of the solution space, including unexpected but valid designs that could inspire human creators.
• The integration of physics simulation with interactive design enabled real-time feedback, supporting incremental learning and co-evolution of human and AI contributions.