What the study found
The study found that large language models (LLMs) could be used to assign tasks across multiple robots when each robot had different room-wise object presence probabilities, meaning estimates of which objects were likely to be in each room. In the reported experiments, the proposed method achieved 47 successful assignments out of 50.
Why the authors say this matters
The authors suggest this approach is relevant for instructions that require searching for multiple objects or handling context-dependent commands, such as requests that are not fully specific. They conclude that their framework can support task decomposition, assignment, sequential planning, and execution for such instructions.
What the researchers tested
The researchers inferred room-wise object presence probabilities using Bayesian inference with a spatial concept model. They converted those inference results into prompts for LLMs, and they designed a few-shot prompting strategy to help the LLM infer required objects from ambiguous commands and break them into subtasks.
What worked and what didn't
The proposed method outperformed the comparison methods reported in the abstract: random assignment had 28 successful assignments out of 50, and commonsense-based assignment had 26 out of 50. The authors also report qualitative evaluation with two actual mobile manipulators, which showed the framework could handle underspecified instructions such as "Get ready for a field trip."
What to keep in mind
The abstract does not describe detailed failure cases, statistical tests, or broader limits of the approach. The reported evaluation includes both a 50-trial assignment experiment and a qualitative test with two mobile manipulators, so the available summary gives only a limited view of performance.
Key points
- LLMs were used to assign tasks in a multi-robot object-retrieval setting.
- Room-wise object presence probabilities were inferred with Bayesian inference and a spatial concept model.
- A few-shot prompting strategy helped the LLM decompose ambiguous commands into subtasks.
- The proposed method achieved 47/50 successful assignments.
- It outperformed random assignment (28/50) and commonsense-based assignment (26/50).
- Qualitative tests with two mobile manipulators handled an underspecified instruction.
Disclosure
- Research title:
- LLM-based task planning improved multi-robot object assignments
- Authors:
- Kento Murata, Shoichi Hasegawa, Tomochika Ishikawa, Yoshinobu Hagiwara, Akira Taniguchi, Lotfi El Hafi, Tadahiro Taniguchi
- Institutions:
- Kyoto College of Graduate Studies for Informatics, Kyoto University, Kyushu Institute of Technology, Ritsumeikan University, Ritsumeikan University, Ritsumeikan University, Ritsumeikan University, Ritsumeikan University, Ritsumeikan University, Ritsumeikan University, Soka University
- Publication date:
- 2026-04-27
- OpenAlex record:
- View
- Image credit:
- Serg (Sergey Kornienko (?)), Wikimedia Commons, GPL
Get the weekly research newsletter
Stay current with scholarly research without reading academic papers — one filtered digest, every Friday.