Tsinghua, Infinigence and Zhengxing Innovation Open-Source RPent as GPT-6 Astra Enters Robot Testing

QbitAI reported that the system reached a 92.6% task success rate on the LIBERO-PRO benchmark and improved end-to-end task completion speed by more than seven times. The work was developed by the core team behind RLinf, a large-scale embodied reinforcement learning framework, and the code is available on GitHub. The report said GPT-6 Astra has begun testing on real robot operations, giving urgency to the question of whether general large models can serve as robot brains.

In demonstration videos, RPent had robots pour steel balls into a bowl, use two arms to wipe a plate, move occluders and find a spoon hidden under a bowl. These tasks were not handled by separate policies trained for each scenario, QbitAI said. Instead, the agent observed the environment, selected capabilities and adjusted actions as conditions changed.

In one example, when the task changed to placing clean tableware into a cardboard box, a frozen VLA continued to place a blue plate into a metal basket as it had learned. RPent first observed the environment and then chose a placement location based on the new goal. When visual localization drifted, it checked the result, ruled out the wrong target and re-localized the plate in its gripper.

In another task, the robot read brand labels on bottles and bags and placed drinks into corresponding bags. After grasping a bottle, it made a small trial lift to confirm a stable grip; when the original motion path was unavailable, it adjusted its wrist posture. For a spoon blocked by bowls, the robot moved the bowls first so the target became visible and reachable. The report described this as a complete observe-judge-execute-correct loop rather than a prewritten action sequence.

RPent also reused existing capabilities. It recombined a pick-and-place container skill for pouring steel balls, and combined grasping with dual-arm coordination and sustained contact for holding, wiping and storing plates. After successful exploration, RPent stored the verified task logic as a Task Card. On later runs, Flash Mode could reuse that workflow and make only necessary adjustments based on current visual state, avoiding repeated calls to the large model at every step and reducing latency, QbitAI reported.

The design separates embodied intelligence into layers rather than relying on one model to handle everything. A general large model handles task understanding and planning; VLA, WAM and other expert models handle high-precision action execution; a memory system stores experience; and the framework connects models, tools and the real environment. The algorithm builds on the team's Harness VLA work published in July.

RPent uses four layers: user, intelligence, interface and environment. The user layer offers a command-line interface and optional web dashboard to issue tasks and inspect visual input, reasoning and action trajectories. The intelligence layer contains an Agentic Planner and Action Primitives, with an Agentic Loop for planning, execution, feedback and memory updates. The interface layer converts decisions into robot-executable instructions. The environment layer supports simulators including LIBERO-PRO, RoboCasa, RoboTwin and RoboDojo, as well as real devices including Franka, dual-arm Franka, YAM and SO101.

The project uses three interface layers: MCP to describe callable capabilities, RPC to connect tools, model services and robot environments, and MHS to provide unified read-write and control abstractions for broader physical devices. Memory is organized into three layers with records for scope, type, credibility and supporting evidence. New experience must be verified before its credibility rises, and conflicting records are retained and isolated. Memory stores Recipes, transferable task solutions rather than fixed coordinates. In a LIBERO-Pro Goal experiment, adding memory raised success in a position-swap task from 31.0% to 87.0%. RPent also supports distributing memory assets so experience from one agent can be shared with others.