Abstract
Effective coordination in decentralized robot teams is often hindered by the "circular reasoning" trap, where agents must recursively model each other's beliefs. This paper proposes a framework that breaks this cycle by assigning specific Speaker and Listener roles to agents. Through these roles, robots communicate implicitly via their physical actions, allowing for seamless collaboration without explicit digital messaging. We demonstrate that dynamic role alternation allows decentralized teams to achieve performance levels comparable to centralized systems in both simulated environments and real-world robotic tasks.
Introduction: The Coordination Challenge
In decentralized multi-robot systems, agents must coordinate their actions based on local observations. Consider two robots transporting a large object; if one robot detects an obstacle, it must convey this information to its partner to ensure a safe trajectory. While explicit wireless communication is a common solution, it can be unreliable or unavailable in many environments.
Humans often solve this problem through implicit communication—inferring a partner's intent or hidden knowledge from their physical movements. However, for robots, interpreting actions is computationally difficult. An agent's movement could be motivated by goal-seeking (exploitation), information sharing (signaling), or environmental exploration. This ambiguity leads to a recursive modeling problem: "I act based on what I think you think I think."
Proposed Methodology: Speaker and Listener Roles
To simplify the inference process, this paper introduces a role-based framework that structures how agents interact and interpret one another:
- The Speaker Role: The Speaker acts purely based on its own local observations and the shared goal. It does not attempt to model the partner's knowledge, instead "exploiting" its current information to move rationally toward the objective.
- The Listener Role: The Listener observes the Speaker’s actions and assumes the Speaker is acting rationally relative to its own view of the world. By "inverting" the Speaker's policy, the Listener can infer hidden state variables, such as the location of obstacles seen only by the Speaker.
- Dynamic Role Alternation: A critical theoretical finding of this work is that static roles can lead to system instability. For a team to successfully navigate complex environments, agents must rapidly alternate roles. This ensures that information from both agents' local sensors is integrated into the team's collective behavior.
The framework is implemented using Artificial Potential Fields (APF). In this model, goals exert an attractive force while obstacles exert a repulsive force. The Listener identifies discrepancies between the Speaker's actual motion and the motion expected from a purely goal-seeking agent, allowing it to calculate the position of inferred repulsive forces (obstacles).
Experimental Evaluation
The proposed framework was evaluated in two distinct settings:
- Simulated Table-Carrying: A continuous state-action environment where two point-robots navigate a rigid object through 1,000 randomly generated obstacle courses.
- Physical Robot Experiment: A real-world setup featuring two Franka Emika Panda robot arms tasked with moving a rod across a table while avoiding obstacles that are occluded from one of the robots' perspectives.
Key Results and Insights
The study yielded several significant findings:
- Performance Parity: In simulations, teams using dynamic role alternation achieved success rates of up to 99.5%, nearly matching the performance of centralized systems with perfect global information.
- Stability through Alternation: Teams with alternating roles significantly outperformed those with fixed "Leader/Follower" roles. Static roles often led to failure because the "Speaker" never incorporated the "Listener's" unique environmental observations.
- Determinism as a Signal: The authors found that for implicit communication to be effective, the Speaker should adopt a more deterministic policy than a centralized controller might. This "cleans up" the behavioral signal, making it easier for the Listener to perform accurate inference.
- Robustness: The role-based approach proved resilient to sensor noise, demonstrating its viability as a fallback mechanism when explicit communication channels are compromised.
Limitations and Future Directions
While effective, the current framework assumes a Decentralized Markov Decision Process (Dec-MDP), where the agents' combined observations cover the entire state. Future work will explore Dec-POMDPs, where parts of the environment remain hidden from the entire team. Additionally, while the theory extends to $N$ agents, physical experiments were limited to two-player teams. Scaling these "Speaker/Listener" dynamics to larger swarms remains a promising area for further research.
Summary updated