Automating the Design of Embodied AgentArchitectures

2026-06-29

AI Summary

Current embodied agent architectures predominantly rely on manual design and lack systematic, automated search methodologies tailored for perception-intensive tasks. This work proposes AgentCanvas—a runtime system coupled with the KDLoop search pipeline—to systematically evaluate the efficacy of architecture search in perception-driven embodied agents for the first time. The approach integrates a closed-loop mechanism of proposal, critique, experimentation, and distillation, enabling editable graph-based programs, simulation rollback for evaluation, and episode-level log analysis. Experiments across four task categories demonstrate that the method automatically generates deployable architectures that achieve targeted improvements in success rates, while also uncovering critical pitfalls in simulation-based optimization—such as data leakage—and highlighting current limitations in the field.

Abstract

Embodied agents are typically built as hand-designed compositions of perception, memory, planning, and action modules. This modularity exposes a large architectural design space, but current systems still rely on researcher intuition to choose where information is stored, how observations are processed, and how model calls are connected. Agent Architecture Search (AAS) automates such design for text-domain agents, but has not been systematically evaluated on perceptual embodied agents through simulator rollouts. We study this transfer. We introduce AgentCanvas, a typed-graph runtime that hosts embodied executors as editable node-and-wire programs with simulator-aware execution and episode-level logs, and KDLoop, a coding-agent search procedure that cycles through proposal, critique, experiment, and distillation, with triggered reflection after stalls. We evaluate three AAS variants across four embodied executors spanning vision-language navigation, embodied question answering, and language-conditioned manipulation. The resulting 3x4 matrix shows that architecture-level search can produce deployable and directional success-rate gains on embodied tasks, while one apparent high-scoring candidate is rejected as leak-bearing. At the same time, the experiments expose constraints that are muted in text-domain AAS: optimization signals can be masked by rollout noise, search can become trapped in local edit basins, and episode-level credit assignment only partially emerges even when detailed logs are available. These results characterize both the promise and the current limits of automated architecture search for embodied agents.

Problem

• Embodied Agents
• Agent Architecture Search
• Automated Design
• Perceptual Agents
• Simulator Rollouts

Innovation

• Agent Architecture Search
• Embodied Agents
• AgentCanvas
• KDLoop
• Automated Design