The RLinf team expanded RPent in September — an open framework that wraps a frozen VLA in an agent with an LLM planner and memory. On the LIBERO-PRO benchmark, π0.5 alone completes 11.0% of tasks. The same model under RPent reaches 92.63%. The code ships under Apache-2.0.
Key takeaways
- 92.63% success (741/800) on LIBERO-PRO with a Codex planner on GPT-6 Astra
- The same frozen π0.5 without RPent: 11.0%
- Apache-2.0 license, 875 stars on GitHub
- September added real-robot support: single-arm and dual-arm Franka
- Non-reasoning mode cuts average execution time by roughly 40%
The frozen model as one tool
VLAs handle precise contact well. Long tasks and shifting conditions, less so. Move the objects or rephrase the instruction and success rates collapse.
RPent does not fine-tune the model. It exposes it as a single retryable "primitive?Primitive: A single, repeatedly callable operation in the agent’s library — here: one call to the frozen VLA, or an analytic operation." instead. Everything else — semantics, navigation, staging before a grasp, release — goes to the planner and a fixed library of analytic operations.
What the leaderboard shows
| Configuration | Success rate |
|---|---|
| π0 | 0.3% |
| X-VLA | 3.8% |
| π0.5 (on its own) | 11.0% |
| RPent + Qwen3.6 27B, no reasoning | 70.63% |
| RPent + Codex on GPT-6 Astra | 92.63% |
Beyond the tabletop simulator, the framework reports 59.20% on RoboCasa365 Target50 and 58.4% on RoboTwin C2R. The July paper claims a 38.6 percentage-point lead over the strongest comparable method on LIBERO-PRO.
Onto real hardware
Until September RPent was mostly a simulation tool. Now there are extensions for Franka arms — single and dual. Supported planners are Claude Code and Codex, with π0.5, RLDX-1 and LingBot-VLA as action primitives.
Why it matters
For two years the field has assumed reliable manipulation runs through bigger models and more training data. RPent points elsewhere: same model, different harness, an eightfold better result. That shifts the centre of gravity from weights to agent architecture — and suggests many existing VLAs are simply being used badly.
What's next
- The leaderboard is public and accepts further configurations, so results can be checked independently
- SO-101 and YAM still sit on the supported-hardware list marked as not ready
- The repository flags the DreamZero world model as an upcoming primitive alongside the VLAs
