PyRUA-Lean

🔥 News

• [Incoming] 🔜 PyRUA-Coder v1.0 (the next release of the trained coder) and PyRUA-RSI, each with its own technical report.
• [2026/11/04] 🚜 PyRUA-DataFarm: paper and code are out. A coding agent takes over the teleoperator's job: from a data request and one calibration episode, it rebuilds the task in simulation and renders photographic, action-labelled multi-view data. On five real-robot tasks, policies trained only on farm data reach 93 successes in 100 trials against 84 for teleoperated data, and 87 against 46 under an unseen tablecloth and coloured light. [Page] [Paper] [Code]
• [2026/11/03] 🏟️ Rank 1 on the RoboColiseum leaderboard: on RoboColiseum, an embodied-AI benchmark jointly initiated by nearly 20 universities, companies and research institutes worldwide, including Hugging Face, NVIDIA and Tsinghua University, PyRUA-Lean with GPT-6.1 Sol and no VLA policy is rank 1 of 26 entries overall (0.8276), with Instruction Following 0.9925, Spatial Reasoning 0.9792, Robustness 0.8492 and General Manipulation 0.4897. [Leaderboard] [Docs]
• [2026/10/30] 🔁 PyRUA-RSI, offline version, open-sourced: one fixed Python program per RoboDojo task, written ahead of time against the same robo API and run with no model call at evaluation time. On RoboDojo eval seed 0 (54 task configurations, 2,100 episodes) it reaches an 82.82% average success rate (Score 87.16); on the held-out seed 2, 80.73% (85.36). [Programs] [Results]. Details in a forthcoming technical report.
• [2026/10/24] 🏔️ #1 on the RoboDojo-Sim leaderboard: PyRUA-Lean is now listed on the official leaderboard with Score 55.26 and a 50.35% average success rate, the first entry above 50%. [Leaderboard]
• [2026/10/21] 🏆 RoboDojo official evaluation: run by the RoboDojo team, PyRUA-Lean reaches a 50.35% average success rate (Score 55.26) on all 42 tasks, rank 1 against the current RoboDojo-Sim leaderboard. [Official results]
• [2026/10/21] 🦾 Real robot: on an AF8 dual-arm robot, GPT-6 Astra writes every Python cell and puts all five tabletop blocks into a bowl, in two independent uncut episodes (41 and 42 cells, about 22 minutes each). [Videos]
• [2026/10/15] 🤗 PyRUA-Lean Coder v0.1-4B is on Hugging Face. [Model]
• [2026/10/13] 📄 arXiv v2 is out as a technical report: it adds RoboDojo and PyRUA-Lean Coder v0.1-4B. [Paper]
• [2026/10/13] 🤖 RoboDojo: without any VLA policy or privileged information, PyRUA-Lean reaches a 51.62% average success rate (Score 58.20), 19.4 points above the best entry of the public leaderboard (snapshot of October 9, 2026). The adapter is open as a pull request to XPolicyLab for RoboDojo's official evaluation (XPolicyLab#168). [Adapter]
• [2026/10/01] 📄 arXiv v1 is out. [Paper]

💡 Introduction

PyRUA (Python for Robot-Use Agents, the robot counterpart of computer-use agents) is a series of projects on robot agents that act through Python code:

• PyRUA-Lean: a token-efficient harness for robot-use agents, with a higher success rate than tool calling, that generalizes across robots and benchmarks; PyRUA-Lean Coder v0.1-4B is trained in it.
  • The framework. The agent composes robot primitives, VLA policies among them, into Python cells that check and retry locally and return only the images and state it asks for. On 700 task instances from LIBERO-PRO, RoboTwin 2.0 and RoboCasa365, with the same GPT-6 Astra planner and the same primitives as tool calling, success goes from 63.1% to 71.7%, with 49% fewer LLM calls and 65% fewer input tokens on the instances both agents solve.
  • RoboDojo. With this library's own motion, gripper and perception primitives and the colour images of the benchmark's default observation, PyRUA-Lean reaches 51.62% average success rate (Score 58.20), first on all five dimensions of the public leaderboard (snapshot of October 9, 2026).
  • PyRUA-Lean Coder v0.1-4B. Qwen3.5-4B trained with GRPO in the PyRUA-Lean framework, on 45 LIBERO-PRO instances, goes from 18.8% to 44.1% on 655 task instances it never saw in training, above the Qwen3.5-27B base model (29.2%).
• PyRUA-DataFarm: a data farm in which a coding agent takes over the teleoperator's job, turning a data request and one calibration episode into real-ready robot data: synchronized photographic camera views labelled with the robot's exact actions. Code on the DataFarm branch.
  • The farm. The agent rebuilds the task in simulation, writing the scene, an expert controller and a success judge; a video model conditioned on real images renders each rollout, and independent judges certify the expert and every rendered episode before admission.
  • Real robot. On five real-robot tasks, π0.5 policies trained only on farm data match policies trained on as many teleoperated episodes (93 vs. 84 successes in 100 trials), and under an unseen tablecloth and coloured light keep 87 successes where the teleoperation-trained policies fall to 46.
  • Cost. Adapting a task is fully automatic and costs $21 of API usage on average, plus $5 for a render contract; every further episode costs $0.80 on average, with no robot or operator.

🚀 Quick start

LIBERO-PRO, RoboTwin 2.0 and RoboCasa365 run on the robot stacks of RPent: install RPent with the robot you want first (the Quick Start goes through it step by step). RoboDojo runs through XPolicyLab and needs no RPent (RoboDojo). Agents run through the Codex CLI 0.155.1, Claude Code, or any OpenAI-compatible endpoint.

git clone https://github.com/DAGroup-PKU/PyRUA-Lean.git && cd PyRUA-Lean
uv pip install -e .       # into the Python environment of your RPent robot stack (or pip)
pyrualean api             # the robot API the agent sees; no simulator needed
npm install -g @openai/[email protected] && codex login      # the agent runtime

# one LIBERO-PRO episode: GPT-6 Astra writes code against robo, at most 40 code cells
export RPENT_ROOT=/path/to/rpent
pyrualean play --arm cells --suite libero_spatial_swap --task 7 --seed 2 \
    --model gpt-6-astra --max-turns 40 --budget-s 7200 --out runs/demo --cuda-device 0
pyrualean summarize runs/demo

The run directory keeps the episode video, every cell the agent ran with its output, the ledger of primitive calls, the token usage and the benchmark's own verdict.

🦾 Robots

Robot	Benchmark	Runs on	Robot and skills
robodojo_xpl_rgb	RoboDojo	XPolicyLab, under RoboDojo's own evaluation client	two ARX-X5 arms; this library's own motion, gripper and perception primitives, SAM3; the three colour images of the default observation; no VLA
libero (default)	LIBERO-PRO	RPent	Franka arm; scripted servos, Pi0.5 VLA, SAM3 perception
robotwin	RoboTwin 2.0	RPent	two arms; LingBot-VLA
robocasa	RoboCasa365	RPent	mobile manipulator; RLDX-1 VLA, navigation
real robot	Demos	AF8 dual-arm robot	two arms, a head camera and two RGB-D wrist cameras; motion planning, gripper and perception primitives with force and torque guards; no VLA; integration coming soon

New robots are added through XPolicyLab (Adding a Robot).

📊 Performance

Code cells against tool calling (GPT-6 Astra, same primitives; calls, tokens and cost per solved episode on the task instances both agents solved):

Benchmark	Success, tool calling	Success, PyRUA-Lean	LLM calls	Input tokens	Cost
LIBERO-PRO	83.0	94.0	21.4 → 8.4	1.23M → 273k	$2.49 → $0.79
RoboTwin 2.0	60.0	69.2	12.2 → 8.9	415k → 245k	$1.00 → $0.66
RC365 atomic	78.9	86.7	10.4 → 5.6	261k → 171k	$0.62 → $0.54
RC365 composite	34.4	39.4	28.4 → 14.9	1.22M → 603k	$2.01 → $1.25
All	63.1	71.7	17.0 → 8.7	788k → 276k	$1.63 → $0.74

RoboDojo (success rate %, the leaderboard's average over its five dimensions; top entries of a snapshot of the public leaderboard taken on October 9, 2026; the official evaluation was run by the RoboDojo team, full results):

Entry	Type	Average	Gen.	Prec.	Long-Hor.	Memory	Open
PyRUA-Lean (GPT-6 Astra), official evaluation	Agent	50.35	43.17	38.25	43.75	68.33	58.25
PyRUA-Lean (GPT-6 Astra), our run	Agent	51.62	45.83	37.25	48.00	63.00	64.00
PyRUA-RSI (fixed programs), our run, seed 0	Code	82.82	72.33	83.50	80.00	95.00	83.25
VPP2	WAM	32.26	29.50	35.17	37.00	55.56	4.08
PhysicalRSI	Agent + VLA	31.38	15.62	32.50	37.33	46.56	24.92
Awomo-0.5	WAM	29.64	23.84	33.50	26.75	41.22	22.92
Simate-beta	VLA	27.96	27.95	26.92	43.42	33.00	8.50

PyRUA-Lean Coder v0.1-4B (success rate %, PyRUA-Lean framework):

Planner	Trained on (45)	LIBERO-PRO other (155)	RoboTwin 2.0 (250)	RC365 atomic (90)	RC365 composite (160)	All unseen (655)
Qwen3.5-4B	37.8	13.5	14.0	52.2	12.5	18.8
Qwen3.5-27B	64.4	38.1	20.8	65.6	13.1	29.2
PyRUA-Lean Coder v0.1-4B	71.1	37.4	52.4	76.7	19.4	44.1

🤗 Acknowledgements

Thanks to the following open-source projects:

• RPent
• XPolicyLab
• RoboDojo
• GenieSim
• SAM 3
• Codex CLI
• miles
• SGLang
• Qwen3.5

🌟 Star History

📖 BibTeX

@article{si2026fewer,
  title={Fewer tokens, better action: GPT-6 Astra robot agents with 14\% higher success rate but 65\% fewer tokens},
  author={Si, Ruiyang and Bi, Jianxin and Yang, Shunyu and Ni, Rui and Huang, Wenbo and Wang, Qiang and Jiang, Shulong and Zhang, Yintianrun and Ruan, Feixiang and Zhang, Kewei and Wang, Duomin and Li, Xiuyu and Feng, Haiwen and Dong, Zhen and Zhou, Daquan},
  journal={arXiv preprint arXiv:2610.01939},
  year={2026}
}

@misc{datafarm2026,
  title  = {PyRUA-DataFarm: LLM Agent as a Cost-Efficient Teleoperator for Real-Ready Robot Data},
  author = {Zhang, Kewei and Ma, Juncheng and Lv, Guodong and Wang, Duomin and Ruan, Feixiang and Si, Ruiyang and
            Zhang, Chuanrui and Chen, Yuxuan and Chen, Hao and Liu, Yichen and Zhang, Yintianrun and Deng, Yufan and
            Xie, Enze and Zhou, Daquan},
  year   = {2026},
  eprint = {2611.05320},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2611.05320}
}