EmboMatrix: A Scalable Training-Ground for Embodied Decision-Making

TL;DR

This work introduces EmboMatrix, a scalable training ground designed to empower large language models with embodied decision-making by combining a multi-agent data factory, a high-throughput scalable simulation backend, and a hierarchical reward framework. The approach yields EmboBrain, an LLM whose embodied capabilities emerge through extensive interactive training, with EmboBrain-7B outperforming substantially larger baselines on challenging embodied benchmarks. Key innovations include a pre-cached physics interface for fast, outcome-based simulation, a distributed simulator network for massively parallel rollouts, and a three-tier reward system that guides learning from format correctness to goal-oriented success. The results demonstrate that environment-grounded, interactive learning can dramatically enhance embodied generalization and efficiency, offering a practical path toward truly intelligent embodied agents.

Abstract

Embodied decision-making enables agents to translate high-level goals into executable actions through continuous interactions within the physical world, forming a cornerstone of general-purpose embodied intelligence. Large language models (LLMs), with their general decision-making capabilities, offer a promising path to realize this potential; however, LLMs trained solely on language lack exposure to physical environments, limiting their true embodied understanding. To bridge this gap, we propose the concept of a training ground: a comprehensive infrastructure that provides task and scene simulation, embodied interaction, and feedback signals, offering a one-stop solution for LLM acquire genuine embodied decision-making skills. In this work, we present EmboMatrix, the first training ground of its kind, providing massive and diverse tasks with efficient simulation and precise rewards. EmboMatrix incorporates a series of novel techniques: a multi-agent data engine for large-scale task and scene generation, a distributed heterogeneous-hardware system for scalable simulation, and a multi-level reward architecture for precise supervision. Leveraging EmboMatrix, we cultivate EmboBrain, an LLM whose embodied decision-making abilities emerge from extensive embodied interactions. Experiments show that EmboBrain-7B surpasses the 671B DeepSeek-R1 baseline by 9.5\% on two challenging embodied decision-making benchmarks, demonstrating the power of interactive, environment-grounded learning for building truly intelligent embodied agents.

Figures (12)

• Figure 1: EmboMatrix $\rightarrow$ EmboBrain: overview and empirical gains.(a) Overall idea.EmboMatrix composes an input prompt from task goals and a library of robot skills, which conditions EmboBrain to generate executable rollouts (e.g., moveto, pickup, open, toggle_on). Execution in the environment returns a reward, closing the loop and enabling prompt adaptation and policy learning. (b) Effect. On the Embodied Agent Interface benchmark, EmboMatrix-augmented training boosts a 7B base model by +58.8 percentage points in success rate, outperforming domain-specialized baselines and even much larger LLMs.
• Figure 2: Overview of the EmboMatrix training pipeline.
• Figure 2: Multi-level layout tree significantly improves the quality of scene generation.
• Figure 3: Pipeline of Multi-Agent-Driven Automated Data Factory.
• Figure 3: Abalation of efficiency. Latency measures the average time to simulate a scenario. We reduces this overhead by over 50×.