MR-STORM: Scalable Multi-Arm Control By Distributed MPC
The paper introduces MR-STORM, a distributed sampling-based MPC framework that utilizes massively parallel GPU sampling and dynamic priority schemes to enable scalable, collision-free multi-arm manipulation in complex environments, demonstrating superior performance in both simulation and physical hardware.
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
In the bustling world of modern robotics, a persistent challenge has long held back the potential of machines working side by side. While a single robotic arm can be programmed to move with precision, coordinating several of them in the same tight space is a different story entirely. When multiple arms operate in a shared workspace, they must constantly avoid bumping into one another while trying to reach their own goals. Traditional methods often rely on a central computer to calculate the perfect path for every arm at once, a process that becomes computationally overwhelming as the number of robots grows. Other approaches let each robot plan its own path independently, but this often leads to confusion, where two arms hesitate indefinitely or crash because they cannot agree on who should move first. The field of motion planning seeks to solve this by creating systems that are both fast enough to react to sudden changes and smart enough to negotiate space without a central boss.
Researchers at Ben-Gurion University of the Negev have developed a new system called MR-STORM to tackle this coordination problem. Their approach allows multiple robotic arms to work together in crowded, dynamic environments by sharing their immediate future plans rather than waiting for a central command. Instead of a single computer calculating every move for a team of robots, each arm uses its own powerful processor to generate thousands of possible movements in a fraction of a second. These arms then broadcast their intended paths to their neighbors. By knowing where a neighbor is about to go, an arm can adjust its own trajectory to avoid a collision before it happens. To prevent the robots from getting stuck in a loop of mutual hesitation, the system uses a simple rule: the arm that is closer to its goal gets to move forward, while the other arm yields. This creates a fluid, decentralized dance of movement where safety and efficiency are maintained without a central coordinator.
The researchers tested this system in a simulated environment using four robotic arms, each with six joints, working in a one-meter square area. They created a variety of challenging scenarios, including tasks where the arms had to reach for targets, follow moving objects, and load items into bins, all while navigating around static obstacles and each other. In these tests, the new system outperformed existing methods that relied on central planning or independent decision-making. The centralized approaches often failed to find a solution as the complexity increased, leaving the robots frozen or moving too slowly to be useful. The independent approaches, while faster, resulted in frequent collisions because the robots could not predict each other's moves. The MR-STORM system, however, successfully balanced speed and safety, allowing the arms to complete tasks efficiently while avoiding contact. In the most difficult scenarios, where workspace density was high, the system maintained a high success rate with very few collisions, demonstrating that sharing plans and using a priority rule effectively resolves conflicts.
To ensure the findings were not just a product of computer simulations, the team deployed the system on a physical humanoid robot equipped with two seven-jointed arms. They programmed the robot to move objects between stations in a grid-like pattern, a task that required the arms to cross paths frequently in a confined space. Over ten trials lasting one hundred seconds each, the robot successfully placed items with minimal error and without any significant collisions. The only contact recorded was a single, light touch of a fingertip, likely due to the simplified way the robot's body was modeled for safety checks. This real-world test confirmed that the system could transfer from a digital simulation to physical hardware, maintaining its ability to coordinate multiple limbs in a cluttered, unpredictable environment.
The study suggests that this decentralized approach offers a scalable solution for future robotic systems, particularly in industrial settings where multiple arms must work in close proximity. By leveraging the parallel processing power of modern graphics cards, the system can generate and evaluate thousands of potential paths simultaneously, allowing for rapid reactions to changing conditions. The researchers found that the key to success was not just in calculating paths, but in how the robots communicated their intentions and resolved conflicts through a dynamic priority system. While the system currently relies on short communication delays and assumes obstacles are relatively predictable, the results indicate a significant step forward in making multi-arm manipulation safe and efficient. The work demonstrates that by giving each robot a voice and a simple rule for yielding, a group of machines can achieve a level of coordination that was previously difficult to attain, paving the way for more complex and collaborative robotic workforces.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.