CompetEvo: Towards Morphological Evolution from Competition

TL;DR

CompetEvo tackles body-brain co-optimization in competitive multiagent environments by co-evolving morphology and fighting tactics in adversarial two-player games. It introduces a delta-uniform, PPO-based training framework that co-evolves morphologies and fighting policies, with a morphology encoding for ant, bug, and spider and their evolvable derivatives. The experiments show morph-evolved agents consistently outperform fixed-morph baselines in both symmetric and asymmetric confrontations and exhibit emergent behaviors such as throwing, wrestling, standing, and defending. This work advances embodied AI by demonstrating how evolvable morphology can adapt to adversarial dynamics, enabling more robust and task-specific agent designs.

Abstract

Training an agent to adapt to specific tasks through co-optimization of morphology and control has widely attracted attention. However, whether there exists an optimal configuration and tactics for agents in a multiagent competition scenario is still an issue that is challenging to definitively conclude. In this context, we propose competitive evolution (CompetEvo), which co-evolves agents' designs and tactics in confrontation. We build arenas consisting of three animals and their evolved derivatives, placing agents with different morphologies in direct competition with each other. The results reveal that our method enables agents to evolve a more suitable design and strategy for fighting compared to fixed-morph agents, allowing them to obtain advantages in combat scenarios. Moreover, we demonstrate the amazing and impressive behaviors that emerge when confrontations are conducted under asymmetrical morphs.

Figures (9)

• Figure 1: The key insight of this article: agent in its original morph is at a disadvantage in competitive confrontations with the opponent. However, after undergoing N generations of co-evolution in both morphology and tactics, agent with new morphology and combat tactics can overcome the original opponent in competition. Using spider and ant as an example.
• Figure 2: Morphology encodings of three different agents: ant, bug, and spider. The encoding methods are specific to the legs. We define 20, 30, and 40 parameters to describe their designs, respectively.
• Figure 3: Information flow in morph and tactics co-evolution training. $x$ denotes initial parameters, which is a randomized vector; $s$ and $o$ are states of the agent and observation of the opponent, respectively; $m$ is generated encoded morph, and $a$ is generated actions applied to each actuator during the confrontation.
• Figure 4: Confrontation environments.
• Figure 5: Morph parameter changes of evolvable agents in the training process and their final evolved morphologies. Each line represents a varying parameter. We mark with dashed lines the correspondence between the most noticeable changes in data and the physical profiles. The red dashed line corresponds to the length data of the limbs, while the black dashed line corresponds to the girth attributes of the limbs.