TLDR: ConPoSe is a novel method that combines Large Language Models (LLMs) with local search to enable multiple robots to cooperatively push objects efficiently. It addresses the scalability limitations of traditional analytical approaches for selecting robot-object contact points. ConPoSe demonstrates superior scalability in selection time and maintains high success rates across various object shapes and robot team sizes, outperforming both analytical and pure LLM-based baselines. While effective, the current contact point switching mechanism presents a challenge for future development.
Object transportation in complex environments is a crucial task for robots, from domestic service to warehouse logistics. When objects are too large or heavy for a single robot, multiple robots must work together. One effective strategy is ‘pushing,’ which can be performed by simpler robots without grasping capabilities. However, a significant challenge in cooperative pushing is carefully selecting the right contact points on the object to ensure it moves along a preplanned path.
Traditionally, this contact point selection can be solved analytically, but the complexity grows exponentially with the number of robots and the object’s size, making it impractical for larger systems. Inspired by how humans use common-sense reasoning for such tasks, researchers have developed a novel approach called ConPoSe.
Introducing ConPoSe: LLM-Guided Contact Point Selection
ConPoSe, which stands for LLM-Guided Contact Point Selection for Scalable Cooperative Object Pushing, combines the powerful reasoning abilities of Large Language Models (LLMs) with a local search method. This hybrid approach aims to select suitable contact points for multiple robots pushing an object. The method successfully handles various object shapes, including cuboids, cylinders, and T-shapes.
The core idea behind ConPoSe is to leverage an LLM to propose an initial set of contact points. If this initial suggestion isn’t optimal or feasible (e.g., not aligned with the target pushing direction or lacking sufficient force), a local search algorithm refines it. This local search explores nearby alternatives to find the best possible configuration, ensuring the object is pushed effectively while minimizing unwanted rotation.
The system operates in a closed loop: it plans a global object path, generates candidate contact points on the object’s contour, selects contact points using ConPoSe, allows robots to switch to these assigned points, and then pushes the object. If the object deviates from the path or a new waypoint is reached, the contact points are re-selected.
Key Advantages and Performance
Experiments conducted in a realistic simulator with up to 15 TurtleBot 4 robots demonstrated several key advantages of ConPoSe:
- Scalability: ConPoSe scales significantly better with an increasing number of robots and object sizes compared to purely analytical methods. For larger robot teams (e.g., 10 or more), the analytical approach becomes computationally intractable, taking days to calculate a single configuration, whereas ConPoSe maintains short selection times.
- High Success Rates: ConPoSe consistently achieves high success rates in moving objects to their goals across various scenarios and robot team sizes.
- Efficiency: Compared to a naive LLM-based selection method, ConPoSe requires fewer contact point switches to follow the preplanned path, making the overall pushing process more efficient.
- Versatility: It effectively handles different object shapes, with cylinders proving to be the easiest to transport due to their rotational invariance and consistent pushing directions.
The research also compared different LLMs, finding that GPT-4.1 offered the best balance between fast selection times and high success rates for their application.
Also Read:
- Smart Hints: LLMs Accelerate Reinforcement Learning in Tricky Environments
- Robots Surpass Human Teachers: A New Learning Method for Enhanced Robotic Performance
Challenges and Future Work
While ConPoSe shows promising results, the contact point switching mechanism remains a significant challenge. Robots, especially nonholonomic differential drive robots like the TurtleBots used, experience slippage and can struggle to maintain stable contact points or navigate in highly constrained spaces with other robots nearby. This can lead to collisions and task failures, particularly for complex shapes like T-shapes with concave corners.
Improving this switching algorithm is a central focus for future work. The authors are confident that their results can be replicated with physical robots, with modern AI methods capable of providing the necessary object information (shape, center of mass) that is currently simulated.
For more in-depth information, you can read the full research paper: ConPoSe: LLM-Guided Contact Point Selection for Scalable Cooperative Object Pushing.
