Tsinghua and Infinigence AI Open-Source RPent, an Embodied Agent Infrastructure with 92.6% Success Rate

Agents in the digital world have already demonstrated a way of working: a large model understands the goal, breaks down the task, calls tools, and adjusts the plan based on results. As GPT-6 Astra begins to undergo real robot operation testing, this paradigm is being pushed into the physical world. But the physical world is not code that can be rolled back at any time—the environment keeps changing, states cannot be fully observed, and once an action occurs it is hard to undo; fine operations such as grasping, assembly, and plugging/unplugging also require specialized models and control capabilities.

RPent's starting point is precisely this gap. Jointly initiated by Tsinghua University, Infinigence AI, and Zhengxing Innovation, it is positioned as embodied agent infrastructure, oriented toward the real physical world, connecting the task understanding and planning capabilities of general-purpose large models, the fine manipulation capabilities of expert models such as VLA, and memory, tools, and robot interfaces, forming a closed loop from perception, decision-making, and execution to feedback and error correction.

What needs to be clarified first is the nature of the evidence: the performance figures, architecture descriptions, and task demonstrations cited in this article all come from QbitAI's report on the RPent open-source project. RPent has indeed been open-sourced (GitHub repository RLinf/RPent), but the benchmark results and real-machine effects in the report have not yet been independently reproduced by a third party, and should be regarded as source-party claims rather than confirmed facts.

Current embodied intelligence broadly presents two routes. One is pure VLA end-to-end: a single vision-language-action model directly takes in observations and outputs actions; it does fine manipulation well, but once tasks get long, language changes, or the scene is perturbed, it degrades rapidly. The other is a pure large-model Agent (such as work like CAP-X): as this generation of general models like GPT-6 improves its coverage of embodied tasks, large models directly outputting actions can already handle quite a few tasks, but there are still obvious shortcomings in sub-centimeter precision, execution latency, and "getting stronger with use."

RPent's position is not to compromise between these two routes, nor to let a single model handle everything from task understanding to robotic arm control. It splits embodied intelligence into capabilities at different layers: a general-purpose large model is responsible for understanding tasks and formulating plans; expert models such as VLA and WAM are responsible for high-precision action execution; the memory system accumulates experience; and the framework is responsible for connecting models, tools, and the real environment. This forms a closed loop of "observe-plan-execute-feedback-replan."

Behind this design there is a traceable algorithmic origin: the report states it derives from the team's Harness VLA work proposed in July (arXiv:2607.08448). Architecturally, RPent is divided into four layers: the user layer, the intelligence layer, the interface layer, and the environment layer. The user layer provides an interactive CLI similar to Claude Code/Codex and an optional Web Dashboard; the intelligence layer consists of the Agentic Planner and Action Primitive; the interface layer converts agent decisions into robot-executable instructions; and the environment layer is responsible for task execution. The intelligence layer and the environment layer adopt an independent process architecture, supporting independent deployment, migration, and restart of model services, robot control, and simulation environments.