Tsinghua, Infinigen, Zhengxing Innovation Open-Source Embodied Agent Infrastructure RPent

On September 21, Tsinghua University, Infinigen, and Zhengxing Innovation jointly open-sourced RPent, an embodied agent infrastructure. It connects general large-model task planning, VLA expert models for fine-grained manipulation, memory, tools, and robot interfaces to form a closed loop of perception, decision, execution, and feedback correction, enabling embodied agents to adapt to environmental changes and continuously learn through interaction.

On the LIBERO-PRO benchmark, RPent achieved a 92.6% task success rate and improved end-to-end task completion speed by more than 7 times. The release also included a series of real-robot demonstrations showing open-ended tasks that do not require a dedicated policy for each scenario. For example, when placing clean dishes into a cardboard box, the agent observes the environment, selects a placement position, and re-checks and re-localizes the item when visual positioning deviates. When sorting drinks by reading brand labels on bottles and bags, it briefly lifts the bottle to confirm a stable grip and adjusts its wrist posture if the original path is blocked. When a spoon is occluded by a bowl, it moves the bowl away before grasping.

The robot also demonstrated recombining a pick-and-place skill to pour steel balls, and combining grasping with dual-arm coordination to hold, wipe, and tidy objects. RPent stores verified task logic as reusable task cards to reduce latency in future executions. The project explores whether robots can reorganize existing skills to complete tasks outside their training distribution, representing the next layer of infrastructure for large models operating in the physical world.