RPent open-sourced: 92.6% on LIBERO-PRO and 7x faster embodied agent execution
量子位 · wechat · 2026-09-21
RPent, an embodied-agent infrastructure open-sourced by a team from Tsinghua, Wuxin Wuqiong and Zhengxing Innovation, connects LLM planning with VLA expert models, memory and robot interfaces into a closed loop for real-world tasks.
- Modular architecture: LLM handles planning, VLA/WAM models handle fine manipulation, a three-layer memory system deposits validated experience (disturbance-task success rose from 31.0% to 87.0%)
- Unified MCP/RPC/MHS interfaces decouple decisions from execution; supports LIBERO/RoboCasa sims and Franka, YAM robots
- FlashMode with TaskCards compresses verified task flows into reusable cards: on LIBERO Object, execution time dropped from 283.6s to 40.9s (7x) with only a 3.5-point success drop
- RPent/GPT-6 Astra scores 92.63% on LIBERO-PRO, 10.23 points above second place; based on the team's July HarnessVLA work
Code: github.com/RLinf/RPent
More from Embodied
- UBTECH sells 400 U1 Pro humanoids to Yongda Auto in ~$10M deal — CyberRobooo · 2026-11-12
- Baidu World 2026 to showcase Apollo Go robotaxi rides and Apollo World physical AI model — Baidu_Inc · 2026-11-12
- NVIDIA's open-source Alpamayo 2 Edge VLA driving model lands in Autoware — drmapavone · 2026-11-12
- A closer look at Astra's robot gripper design (and a joke at its expense) — _Stocko_ · 2026-11-12
- Robot startup passes first presale milestone with 10+ units claimed — R4nsack · 2026-11-12
- Runway launches Praxis-1, a 27B MoE world action model, with a 73ms Fast variant — ChongZzZhang · 2026-11-12