Tsinghua University, Infinigence AI and Zhengxing Innovation Open-Source RPent Embodied-Agent Platform

leaderobot.com reported on September 21, 2026, that Tsinghua University, Infinigence AI and Zhengxing Innovation jointly open-sourced RPent, an embodied-agent platform. The article says RPent connects general-purpose large models, specialized models such as VLA, memory, tools and robot interfaces to form a closed loop of perception, decision-making, execution and feedback. It also reported that RPent achieved a 92.6% task success rate on LIBERO-PRO and increased end-to-end task completion speed by more than sevenfold.

RPent, jointly open-sourced by Tsinghua University, Infinigence AI and Zhengxing Innovation, is notable for making public not only the part that makes robots move, but also mechanisms for reusing successful procedures. The focus is shifting from research demos to how to ensure reproducibility and reusability in real-world operations, showing that one-off performance metrics alone are not enough.

What is new about RPent is not the addition of individual motion models, but the release of a unified platform that connects general-purpose large models, VLA-based execution, memory, tools and robot interfaces, and includes an operational system for reusing exploration results as Task Cards. Rather than the grasping performance of an individual model or the polish of a demo, the design emphasizes carrying successful procedures forward for future use. This can be read as a shift from research-stage robot control toward a reusable execution platform. The figures of 92.6% on LIBERO-PRO and more than sevenfold end-to-end speed also show that evaluation is expanding beyond movement alone to the entire task-completion process. A related article from us, Infinigence AI, Tsinghua University and Shanghai Jiao Tong University release APXInf, an embodied edge-inference engine, pointed to efforts to make inference on the robot itself more lightweight. That article focused on reducing edge-inference latency, while this release differs in that it designs a closed loop encompassing not only inference but also memory and mechanisms for rerunning tasks. In other words, lighter computing infrastructure and the accumulation and reuse of action experience are beginning to connect within the same execution system, rather than remaining separate themes. The industry-wide implication is that the competitive focus in robotics could shift from high-performance one-off models to execution frameworks that can be used repeatedly in the field. Mechanisms such as Task Cards and Flash Mode point toward retaining verified procedures rather than rebuilding all actions every time a model is updated, potentially affecting the design of training data, evaluation criteria and operational maintenance. However, the article presents demos and benchmarks; uptime in real-world environments, recovery from failures and the range of tasks to which the system can generalize remain unverified. See the original source for details.

For Japan’s robotics industry, the issue is likely to be not only the performance of individual machines, but also how to build layers for execution frameworks and experience reuse. In applications where procedures are difficult to standardize in the field, reusable designs such as Task Cards could affect how control software and training infrastructure are separated or integrated.