EmboCoach-Bench: Benchmarking AI Agents on Developing Embodied Robots

TL;DR

EmboCoach-Bench presents a first-of-its-kind benchmark to evaluate Large Language Models as autonomous engineers in embodied AI, moving beyond static code generation to a dynamic, environment-aware, repository-level development loop. By formalizing each task as a triplet T = (D_prd, P_sys, C_env) and providing an integrated Agent-Environment Interface, the framework enables Draft-Debug-Improve cycles that leverage environment feedback to co-evolve policies and architectures across 32 tasks and four simulators. Empirical results show consistent, substantial gains over human baselines and single-shot baselines, with agents recovering from pathological failures and achieving high or near-perfect success in several tasks. The work demonstrates a scalable path toward self-evolving embodied intelligence, reducing reliance on manual engineering and moving toward autonomous, compute-powered robotic development that can generalize across platforms and modalities, thereby accelerating real-world deployment of embodied AI systems.

Abstract

The field of Embodied AI is witnessing a rapid evolution toward general-purpose robotic systems, fueled by high-fidelity simulation and large-scale data collection. However, this scaling capability remains severely bottlenecked by a reliance on labor-intensive manual oversight from intricate reward shaping to hyperparameter tuning across heterogeneous backends. Inspired by LLMs' success in software automation and science discovery, we introduce \textsc{EmboCoach-Bench}, a benchmark evaluating the capacity of LLM agents to autonomously engineer embodied policies. Spanning 32 expert-curated RL and IL tasks, our framework posits executable code as the universal interface. We move beyond static generation to assess a dynamic closed-loop workflow, where agents leverage environment feedback to iteratively draft, debug, and optimize solutions, spanning improvements from physics-informed reward design to policy architectures such as diffusion policies. Extensive evaluations yield three critical insights: (1) autonomous agents can qualitatively surpass human-engineered baselines by 26.5\% in average success rate; (2) agentic workflow with environment feedback effectively strengthens policy development and substantially narrows the performance gap between open-source and proprietary models; and (3) agents exhibit self-correction capabilities for pathological engineering cases, successfully resurrecting task performance from near-total failures through iterative simulation-in-the-loop debugging. Ultimately, this work establishes a foundation for self-evolving embodied intelligence, accelerating the paradigm shift from labor-intensive manual tuning to scalable, autonomous engineering in embodied AI field.

Figures (13)

• Figure 1: Overview of the EmboCoach-Bench Framework. This diagram illustrates the end-to-end workflow for evaluating AI engineering agents in embodied AI.
• Figure 2: Formalizing the Engineering Task. An instantiation of the task tuple $\mathcal{T}$. The Specification ($\mathcal{D}_{prd}$) strictly enforces constraints (e.g., 4-hour limit) and offers scaffolding. Crucially, the Substrate ($\mathcal{C}_{env}$) requires the agent to navigate and modify four distinct modules (Policy, Config, Workspace, Dataset) to achieve the objective, simulating repository-level complexity.
• Figure 3: The composition of the EmboCoach-Bench benchmark. The structure is organized into three hierarchical layers spanning foundation, breadth, and depth. (Bottom) The Foundation: Built upon four widely adopted, high-fidelity simulators ensuring reproducibility. (Middle) The Task Spectrum: A curated suite of 32 diverse tasks covering both RL and IL paradigms, ranging from basic rigid-body interactions to complex fine motor skills and articulated object manipulation. (Top) The Skill Depth: Probes advanced engineering proficiency, focusing on critical challenges such as reward engineering in RL and implementing SOTA architectures like Diffusion Policies, ACT, and VLA models in IL.
• Figure 4: The Task Formalization Framework. Each benchmark task is modeled as a tripartite tuple $\mathcal{T} = (\mathcal{D}_{prd}, \mathcal{P}_{sys}, \mathcal{C}_{env})$. (Left) Semantic Specification ($\mathcal{D}_{prd}$): A structured PRD acting as the guiding objective, defining personas, constraints (e.g., resource budgets), and domain scaffolding. (Middle) Operational Interface ($\mathcal{P}_{sys}$): A system-injected protocol defining the API schemas for the FileEditor, Terminal, and TaskTracker. (Right) Development Substrate ($\mathcal{C}_{env}$): The target environment, encapsulating the simulator infrastructure (Docker/K8s) and the human baseline codebase that the agent must audit and optimize.
• Figure 5: Average success rate on EmboCoach-Bench. We compare various state-of-the-art LLMs. The comparison highlights the significant impact of the agentic framework ("agentic", hatched bars) versus a standard direct prompting baseline ("w/o agentic", solid bars). The horizontal line indicates the human baseline performance.