Swarm-Coordination-Watch-Collision-Dataset

2,224 state samples from simulated 2-agent grid navigation, labeled with whether a collision occurs within the next 3 steps, for training early-warning collision prediction models in multi-robot systems.

About Dataset

This dataset contains 2,224 individual state samples drawn from 150 simulated episodes of two agents navigating a shared grid, each trying to reach its own goal without colliding with the other. Agents use a mix of cautious (collision-avoiding) and reckless (goal-only) heuristic policies. Each sample records the current state (distance between agents, whether they're heading toward each other, each agent's distance to its own goal) and whether an actual collision occurs within the next 3 timesteps. Coordination failures (collisions, deadlocks) are a real, practically important problem for multi-robot systems - warehouse robot fleets, autonomous vehicle coordination, and drone swarms all face this challenge (Lowe et al., 'Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments', NeurIPS 2017). This dataset targets early detection - predicting a collision before it happens, not just recording it afterward.

Purpose of Dataset

To Support Development Of Early-warning Systems For Multi-robot Coordination Failures, Enabling Real-time Intervention (Yielding, Replanning) Before A Collision Occurs Rather Than Only Handling It After The Fact.