Less is More: Robust Zero-Communication 3D Pursuit-Evasion via Representational Parsimony
New research shows simpler AI agents with 40% fewer observation channels outperform complex, communication-dependent systems.
A research team from unnamed institutions has published a breakthrough paper on arXiv titled 'Less is More: Robust Zero-Communication 3D Pursuit-Evasion via Representational Parsimony.' The work tackles the complex problem of coordinating multiple autonomous drones (pursuers) to capture a target (evader) in dense, three-dimensional voxel environments without any communication between agents. This scenario is critical for real-world applications where communication latency, jamming, or bandwidth constraints render traditional multi-agent systems ineffective.
The core innovation is a principle called 'representational parsimony,' which deliberately simplifies what each AI agent perceives. The team reduced the observation space for each pursuing drone from 83 dimensions to just 50 by removing channels that relayed information about teammate states. To compensate for this lack of direct coordination data, they developed a novel training mechanism called Contribution-Gated Credit Assignment (CGCA), which helps agents learn cooperative behaviors by rewarding actions that contribute to the team's local progress.
In rigorous simulations, a team of 4 pursuers using this 'parsimonious' method achieved a 75.3% success rate (±0.091) with a collision rate of 22.3%, outperforming a baseline system with full 83-D observations (72.1% success, 25.3% collisions). The system demonstrated remarkable robustness, maintaining performance under stress tests involving added sensor noise, control delays, and speed mismatches. It also showed strong 'zero-shot' transfer, achieving about 61% success when deployed without retraining in a new, complex 'urban-canyon' map environment.
The results challenge the prevailing assumption in Multi-Agent Reinforcement Learning that richer inter-agent information sharing always leads to better coordination. Instead, the research suggests that strategically limiting observation can prevent 'compounding error cascades' and create more robust, decentralized AI systems capable of operating in communication-denied environments, a key requirement for future military, security, and search-and-rescue drone swarms.
- Achieved 75.3% success rate with 4 AI pursuers vs. 1 evader in 3D cluttered environments using zero communication.
- Reduced observation dimensions by 40% (83-D to 50-D) and introduced CGCA for decentralized credit assignment, improving robustness.
- Demonstrated resilient performance in stress tests and 61% zero-shot transfer success to unseen 'urban-canyon' maps.
Why It Matters
Enables robust, decentralized drone swarms for defense and rescue missions where communication is unreliable or adversarial.