AI xPress

AI and computer science research.

Back to Foragentis

Language & LLMs

Program-like prompts helped a large language model achieve state-of-the-art success rates on VirtualHome household robot tasks, a 2023 study reports

Language & LLMs October 31, 2026

A classroom wall covered with paper speech bubbles

In a 2023 paper, researchers from the University of Southern California and Nvidia describe a prompting method for robot task planning. They gave a large language model, a model trained on large amounts of text, program-style descriptions of the actions and objects available in a setting. They also included example programs that could be executed. The authors report state-of-the-art success rates on household tasks in VirtualHome. They also ran the method on a physical robot arm for tabletop tasks, but the abstract does not report a measured outcome for those physical trials.

What problem does the method address?

Robot task planning often requires a lot of hand-written knowledge about the world where a robot acts. Some earlier language-model approaches require listing every possible next action and scoring each one. Others produce free-form text that may include actions the robot cannot perform in its current setting. This paper proposes a method that avoids both limits by giving the model program-like information about the environment.

How did the prompt work?

The prompt gave the model program-style descriptions of the available actions and objects, along with example programs that could be executed. The researchers also changed parts of the prompt and the constraints on how the model produced text in ablation experiments, which test the effect of removing or altering parts of a method. They report that these tests produced concrete recommendations about prompt structure, though the abstract does not list the recommendations.

What did the researchers find?

The authors report that the method achieved state of the art success rates on household tasks in VirtualHome. The abstract does not include a specific success-rate figure. The team also ran the method on a physical robot arm for tabletop tasks, but the abstract does not state a measured result from that setting.

Frequently asked questions

What is VirtualHome?

The abstract names VirtualHome in connection with the household tasks on which the researchers measured success rates. It does not describe what VirtualHome is.

Did the researchers test the method outside simulation?

Yes, they did. The abstract says the researchers also ran the method on a physical robot arm for tabletop tasks. It does not report a measured success rate for those physical trials.

Does the abstract report exact success numbers?

No. It reports success rates only in general terms for VirtualHome household tasks.

Limits and what's next

The measured success-rate claim comes from VirtualHome household tasks. The physical robot test involved tabletop tasks on one robot arm, and the abstract does not report a measured outcome for that part. The paper used ablation experiments to compare design choices, but the abstract does not list the exact conditions changed. It also does not give a specific success-rate figure. The abstract reports no independent replication or external validation. The result is one study's report from a virtual household setting and a demonstration on a physical tabletop task.

Source: Ishika Singh et al., “ProgPrompt: Generating Situated Robot Task Plans using Large Language Models,” Proceedings - IEEE International Conference on Robotics and Automation/Proceedings, 2023.