PhysicalWorldAgent GitHub

When the WorldIs theTest Suite

Agentic Intelligence for the Physical World

Kaixin Li1 Meng Luo1 Yanlin Li1 Yuchen Tian2 Xiang Liu1 Xiangru Jian3 Yan Yang4 Zhiheng Xi5 Yang Zhang1 Ziyang Luo2 Mingfei Cheng6 Hongzhan Lin1†

1NUS 2HKBU 3UWaterloo 4ANU 5Fudan 6SMU † Corresponding author

The test suite does not ship with the code

Coding agents now write programs for robots, laboratories, and industrial controllers. Whether a program works can only be found out by running it on a robot, in a simulator, or through the policy it trains.

Writing programs is getting cheap. Trustworthy evidence about them is not.

Fig. 1Same program, different test suite.In software, the test suite usually ships with the code and returns a cheap, repeatable result. Once the program acts in the physical world, someone must build the judge, a simulator can stand in for the robot, and every physical test has to be paid for in execution.
Read the full abstract

Coding agents built to repair software now write, test, and repair programs for robots, laboratories, and industrial controllers. Unlike in software engineering, no test suite comes with the code. Whether a program works must be established by running it on a robot, in a physics simulator, or through the policies it helps to train, and such evidence is costly to obtain and can itself be wrong. As these agents take on more physical tasks, how their programs are judged matters as much as how they are written. This paper examines this emerging area, agentic intelligence for the physical world, by treating the world as the program's test suite.

We organize the discussion around three questions: how agents write programs for the physical world, how these programs are tested, and how agents learn from the results. We first trace how language models for robots and software engineering agents came together, ask which software testing practices carry over to physical tests, and describe each program by where it sits relative to the actuator and by what judges it. For writing, we cover programs that act as robot policies and programs that support policy learning by producing rewards, tasks, and training worlds. For testing, we examine what physical trials cost, what traces, learned judges, simulators, and formal checks can establish in their place, and what safety requires beyond task success. For learning, we discuss how tested programs are retained as experience, in skill libraries and in model weights. Finally, we turn to benchmarks and evaluation practice and outline open challenges. Writing programs is becoming cheap, while trustworthy evidence about what they do remains costly. Progress will therefore depend on a physical counterpart of continuous integration: verification cheap enough to run often, reliable enough to trust under search, and maintained alongside the experience it validates.

Fig. 4Four phases of agentic intelligence for the physical world.Language models for robots moved from choosing skills to writing programs, and coding agents from software engineering joined this line in 2026.

Where the loop breaks

The agent's grasp aimed at the pot, not its handle. It rewrites the routine.

A program can only use the primitives it is given.

Fig. 2When the world is the test suite.The kitchen robot writes a program to put a moka pot on the stove ①, runs it ②, has the outcome judged ③, and repairs the program or keeps what passes ④.
Fig. 5Where software assumptions need support in the physical world.A physical task needs a success criterion ①, an account of trial-to-trial variation ②, a reset ③, a budget for costly tests ④, limits on unsafe commands ⑤, and a statement of what the agent can observe ⑥.

Programs that act, and programs that teach

Some programs drive the robot or orchestrate learned skills. Others write the rewards, worlds and demonstrations a policy learns from.

Fig. 6Programs as robot policies: four questions.What a program can call ①, when it is revised ②, how it coordinates learned skills ③, and when it outperforms an end-to-end policy ④.
Fig. 7What the code-versus-VLA comparisons on LIBERO-PRO cover.Gray bars: OpenVLA, π0, π0.5 under five perturbations. Colored bars: program-based methods, reported only under position and task perturbations (shaded).
Fig. 8Programs for policy learning.A reward program shapes what a policy learns ①; programs that build environments, demonstrations, bodies, and assets shape where it learns ②–④.

Evidence is costly, and it can be wrong

Fig. 9How an agent learns that a program is wrong.Traces ①, judges ②, simulation ③, formal specifications ④, release checks when trials are restricted ⑤, and safety beyond success ⑥.
Fig. 10Evaluation paths for different kinds of programs.From checking code to training and testing the behavior it induces, with the budgets reported per execution, per candidate, or per search.

What gets kept

Accepted programs become experience, in skill libraries and in model weights. A library keeps whatever its acceptance test lets through.

Fig. 11What happens to a program that passes its checks.It is stored ①, often after an acceptance test ②; the library must be maintained ③, may transfer ④, and can train a model's weights ⑤.
Fig. 12What a growing skill library buys, and where it hurts.ASPIRE on LIBERO-PRO Long: average success rises with the library, while the stove and moka pot task falls from 0.68 to 0.26 between 50 and 90 source tasks.

Where next

Fig. 13Four challenges and future directions.Program representation ①, reliable verification ②, skill maintenance ③, and reproducible evaluation ④.
Fig. 3Organization of the article and representative work.Each panel is one part of the article and lists representative work discussed there.

If the world is the test suite, its tests must be tested too.

Cost · Variance · Judge error · Judge visibility · Sim-to-real

What most reports leave out.

Cite

@article{li2026world,
  title   = {When the World Is the Test Suite: Agentic Intelligence for the Physical World},
  author  = {Li, Kaixin and Luo, Meng and Li, Yanlin and Tian, Yuchen and Liu, Xiang and Jian, Xiangru and Yang, Yan and Xi, Zhiheng and Zhang, Yang and Luo, Ziyang and Cheng, Mingfei and Lin, Hongzhan},
  journal = {arXiv preprint arXiv:2611.22291},
  year    = {2026},
  url     = {https://arxiv.org/abs/2611.22291}
}