Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer Tokens
Ruiyang Si, Jianxin Bi, Shunyu Yang, Rui Ni, Wenbo Huang, Qiang Wang, Shulong Jiang, Duomin Wang and 4 others
Peking University · National University of Singapore · NVIDIA · Impossible Research
Published October 1, 2026 · on the feed October 2, 2026 · ▲ 47 · arXiv 2610.01939
The abstract
Vision language model (VLM) agents can control robots through visual feedback and action primitives, but repeated model invocations and redundant observations incur substantial token overhead. We introduce PyRUA-Lean, an interactive code-execution framework that couples feedback-driven primitive composition with selective observation: the agent composes classical robot primitives and learned vision-language-action (VLA) policies into Python cells that perform conditional checks and local retries, returning only explicitly requested images and state feedback for replanning. Across 700 simulated task instances from LIBERO-PRO, RoboTwin 2.0, and RoboCasa365, we compare PyRUA-Lean with a tool-calling baseline using the same GPT-6 Astra planner and underlying robot primitives. Under equal LLM-call budgets, PyRUA-Lean increases overall success from 63.1% to 71.7%. On instances solved by both agents, it uses 49% fewer LLM calls and 65% fewer input tokens.
The Brief
2:04A short podcast summary of the paper. Written and voiced by AI.
Transcript
Fewer Tokens, Better Action. GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer Tokens. By Ruiyang Si and eleven others, from Peking University, the National University of Singapore, and two other institutions. Published on the first of October. Its code repository is worked out: about forty-five source files, with a readme and a licence.
Vision language model agents can control robots through visual feedback and action primitives. The cost is in the loop. Repeated model invocations and redundant observations add up to substantial token overhead. Each step that sends the model back a fresh image, and asks it to decide again, is paid for in tokens, whether or not the step needed a decision.
PyRUA-Lean is an interactive code-execution framework built around two habits. Feedback-driven primitive composition, and selective observation. The agent writes code that does more per call, and sees only what it asks to see.
Concretely, the agent composes classical robot primitives and learned vision-language-action policies into Python cells. Those cells perform conditional checks and local retries themselves. What returns for replanning is only the images and state feedback the agent explicitly requested. A check that passes, or a retry that succeeds, can happen inside the cell, without a round trip to the planner.
The evaluation covers seven hundred simulated task instances from LIBERO-PRO, RoboTwin two point zero, and RoboCasa365. The comparison is against a tool-calling baseline that uses the same GPT-6 Astra planner and the same underlying robot primitives. Under equal LLM-call budgets, overall success rises from sixty-three point one percent to seventy-one point seven percent. On instances both agents solve, PyRUA-Lean uses forty-nine percent fewer LLM calls, and sixty-five percent fewer input tokens.
Same planner, same primitives: only the interface between them changed. Moving checks and retries into code, and returning observations on request, is enough to raise success while cutting what the planner has to read.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
The robot works. The bill is the problem: every step sends the model back for another look.
You run a robot through a vision language model that calls one operation at a time: move, look, grasp. After each step the model is invoked again, and several camera images are added to the conversation automatically.
The road ahead
- 1 · The problem ~7 min When a vision language model runs a robot one operation at a time, every step costs a model call and a pile of images.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min The agent writes Python code that chains robot operations, checks and retries them itself, and sends back only what it asks for.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min The agent writes a cell of robot operations, the cell checks and retries, returns only what was asked, and the planner writes the next cell.
- 6 · The experiment ~6 min You’ll guess: On task instances both agents solved, how did the code agent's number of LLM calls compare with the tool-calling agent's?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.