💬 NLP

Leveraging Adaptive Group Negotiation for Heterogeneous Multi-Robot Collaboration with Large Language Models

This paper introduces CLiMRS, an adaptive group negotiation framework that leverages Large Language Models to enable efficient and robust heterogeneous multi-robot collaboration through dynamic subgroup formation and perception-driven discussions.

Original authors: Siqi Song, Xuanbing Xie, Zonglin Li, Yuqiang Li, Shijie Wang, Biqing Qi

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Siqi Song, Xuanbing Xie, Zonglin Li, Yuqiang Li, Shijie Wang, Biqing Qi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to organize a massive, chaotic wedding reception. You have a caterer, a florist, a DJ, and a decorator. If you just shout, "Everyone, get to work!" and walk away, you’ll end up with the DJ playing music while the cake is still in a box, or the florist trying to decorate a table that hasn't even been set yet.

This paper, CLiMRS, is essentially a "Smart Wedding Planner" for a team of different robots.

The Problem: The "Too Many Cooks" Dilemma

Usually, when scientists teach robots to work together, they use "homogeneous" teams—meaning they use a bunch of identical robots that all do the same thing (like a swarm of tiny vacuum cleaners).

But the real world is heterogeneous. In a real factory or a home, you have a "specialist" robot arm that can pick things up, a "heavy-lifter" robot on wheels that can carry boxes, and a "scout" robot that can walk over obstacles. Getting these different "specialists" to talk to each other and coordinate a long, complex task (like building a machine from scratch) is incredibly hard. They often bump into each other, get confused by mistakes, or simply don't know who should do what next.

The Solution: CLiMRS (The Smart Manager)

The researchers created a system called CLiMRS. Instead of one giant "brain" trying to control every single motor of every single robot (which would be overwhelming), they gave each robot its own "mini-brain" (an LLM, like ChatGPT) and then added a layer of Adaptive Group Negotiation.

Here is how it works, using a Construction Site analogy:

1. The General Contractor (Grouping)

Instead of everyone standing around, a "General Contractor" (the Proposal Planner) looks at the blueprints and the workers. It says: "Okay, we need to build a wall. Group A (the bricklayers) go to the corner. Group B (the cement mixers) stay by the truck. Group C (the inspectors) watch the progress." It dynamically creates small "work crews" so robots aren't tripping over each other.

2. The Crew Leads (Planning)

Inside each small crew, there is a "Crew Lead" (the Subgroup Manager). This lead doesn't worry about the whole building; they only worry about their specific task. They talk to their crew members: "Hey, Bricklayer #1, grab that bucket. Bricklayer #2, get ready to catch it."

3. The Workers (Execution)

The robots themselves are the workers. They receive the command, check if they are physically capable of doing it (e.g., "Can I actually reach that brick?"), and then do the work.

4. The "Oops" Factor (Feedback)

This is the most important part. In the real world, things go wrong. A robot might slip, or a box might fall. In old systems, a robot would just fail and stop. In CLiMRS, the robot "talks back." It says: "I tried to pick up the wheel, but it's too heavy/slippery!"

The "Smart Manager" hears this, realizes the plan has failed, and holds a quick meeting: "Okay, the wheel is too heavy for the small robot. Let's re-group. Heavy-Lifter Robot, you take over this task."

Why is this a big deal?

The researchers tested this in a digital "obstacle course" called CLiMBench, where robots had to find parts and assemble a wheeled robot.

The results? Their system was 40% more efficient than previous methods. While other robots were getting stuck in loops or running out of time because they couldn't handle mistakes, CLiMRS was able to "negotiate" its way through errors, re-assigning tasks on the fly until the job was done.

Summary in a Nutshell

CLiMRS turns a group of specialized robots from a confused crowd into a highly organized, self-correcting professional team that can plan, talk, and fix their own mistakes.

Technical Summary: Leveraging Adaptive Group Negotiation for Heterogeneous Multi-Robot Collaboration with LLMs

1. Problem Statement

Multi-robot collaboration in real-world scenarios involves several complex challenges:

Heterogeneity: Robots possess diverse physical capabilities (e.g., manipulation, transportation, bipedal locomotion) that must be coordinated.
Long-Horizon Planning: Tasks like assembly require a long sequence of interdependent actions.
Uncertainty and Errors: Real-world execution is prone to failures (e.g., collisions, improper grasping, or environmental changes) that can derail a static plan.
Scalability vs. Efficiency: Centralized planning for many agents can become computationally expensive and communication-heavy, while purely decentralized approaches often struggle with global coordination and task decomposition.

2. Methodology: The CLiMRS Framework

The authors propose CLiMRS (Cooperative Large-Language-Model-Driven Heterogeneous Multi-Robot System), a framework inspired by human teamwork principles. It utilizes a grouping–planning–execution–feedback loop to manage complexity through adaptive negotiation.

The framework consists of five core modules:

General Proposal Planner (Grouping): An LLM that analyzes the global task, robot capabilities, and environment to dynamically partition agents into task-oriented subgroups. This allows for parallel execution and minimizes interference.
Subgroup Managers (Planning): Multiple LLMs that operate in parallel. Each manager leads a centralized discussion within its assigned subgroup to generate fine-grained, robot-specific commands based on local observations.
Agent Executors (Execution): Each robot is paired with an independent LLM agent. The executor verifies if the command is physically feasible for its specific hardware before issuing the low-level skill command.
Simulation Environment: A physics-based backbone (IsaacGym) that executes actions and provides real-time feedback regarding environmental changes and potential conflicts.
Context Memory: A centralized module that stores dialogue history, agent-level execution feedback (including diagnostic explanations for failures), and environmental observations to inform the next cycle of the loop.

3. Key Contributions

CLiMRS Framework: A novel architecture that leverages the reasoning of LLMs to enable dynamic subgroup formation and multi-level negotiation (global $\rightarrow$ subgroup $\rightarrow$ individual agent).
CLiMBench Benchmark: A new, challenging benchmark for heterogeneous multi-robot collaboration. It features long-horizon assembly tasks (e.g., building a wheeled robot) in a realistic physics simulator, explicitly modeling execution uncertainty and partial observability.
Robustness through Feedback: Unlike previous benchmarks that assume perfect execution, CLiMBench and CLiMRS focus on the ability to recover from execution failures using natural language feedback.

4. Experimental Results

The framework was evaluated on the existing COHERENT benchmark and the new CLiMBench.

Efficiency and Success: In CLiMBench, CLiMRS achieved a 100% success rate, outperforming all baselines. Notably, it reduced the Average Steps (AS) by over 40% compared to the best baseline, demonstrating superior planning efficiency.
Complexity Handling: CLiMRS showed its greatest gains in "trio-type" tasks (requiring three different robot types), proving that its grouping strategy is highly effective for complex coordination.
Ablation Insights: Ablation studies confirmed that dialogue history and execution feedback are indispensable. Removing the grouping stage significantly degraded performance, highlighting the importance of parallelized, subtask-oriented coordination.

5. Significance

This work represents a significant step toward embodied multi-agent intelligence. By moving away from rigid, centralized planning and toward an adaptive, "negotiation-based" approach, the authors demonstrate that LLMs can effectively manage the complexities of heterogeneous robot teams. The ability to recover from physical execution errors through linguistic reasoning provides a robust pathway for deploying multi-robot systems in unpredictable, real-world environments.