Key Takeaway
Train one shared, target-conditioned controller for a single agent and use optimal transport to assign targets — you get better coordination, big speedups in training, and easy scaling to many agents.
ON THIS PAGE
Key Findings
A separation between two problems—assigning which agent should go to which target, and how a single agent should reach a given target—works well for many cooperative tasks. Learn a single-agent controller conditioned on a target and use Orchestrator-Worker Pattern to assign targets across the fleet; this avoids learning in the huge joint state-action space. The resulting family of methods (SALT) reduces travel time in realistic routing tasks and can be trained far more sample-efficiently than standard multi-agent training. The guarantee is exact when dynamics have no randomness and provides a bounded approximation gap under stochastic dynamics. Dynamic Task Routing Pattern
Not sure where to start?Get personalized recommendations
Data Highlights
115–21% reduction in travel time on a realistic southern Manhattan road network versus a shortest-path assignment heuristic.
223–61% average travel time reduction in controlled stochastic grid worlds compared to standard multi-agent baselines.
3Up to 480× fewer environment interactions to reach performance milestones and successful scaling to 128 agents, while a popular joint-training baseline failed by 8 agents within a 2×10^9 interaction budget.
What This Means
Engineers building fleet coordination (delivery, ride-hailing, warehouse robots) will get better coordination without retraining for new fleet sizes. ML engineers and researchers can use any single-agent learner inside this approach to massively reduce training cost and improve sample efficiency, while keeping a simple, modular coordination layer. Agent Registry Pattern
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The approach relies on agents having decoupled dynamics and a terminal cost that can be written as an assignment-style transport cost, so tasks with strong interactions (collisions, coupled constraints) or non-transport objectives are outside the theoretical guarantee. The separation gives an upper bound rather than an exact value under stochastic dynamics, and the gap grows with noise and state dispersion. Success also depends on the single-agent controller generalizing across the target space, which becomes harder when targets are high-dimensional or drawn from very different distributions than those seen in training. State Inconsistency routing
Full Analysis
Separate target assignment from single-agent control: learn a single-agent cost-to-go or policy conditioned on a target, then solve an optimal transport assignment across the population using that learned cost as the transportation cost. The transport step assigns which agent should go to which target, while the single-agent controller tells an assigned agent how to reach its target. Because the single-agent problem is independent of the fleet size, training and representation remain compact and generalize across population sizes. Theoretical results show the fleet-level cost is upper bounded by the optimal transport cost that uses the target-conditioned single-agent cost-to-go; the bound is exact when dynamics are deterministic and bounded in the stochastic case. Practical algorithm SALT (Separation-based Assignment and Learning via optimal Transport) plugs any single-agent learner (Q-learning, PPO, etc.) into the framework and uses Hungarian or entropic transport solvers to assign targets. Empirically, SALT outperforms multi-agent baselines across noise models, reduces travel time by 15–21% on a large Manhattan road network, achieves 23–61% gains in grid worlds, and is up to hundreds of times more sample-efficient, scaling to dozens or hundreds of agents where joint training fails. Model Context Protocol (MCP) Pattern Routing
Explore evaluation patternsSee how to apply these findings
Credibility Assessment:
Includes a recognizable top researcher (Florian Dörfler) known for strong work in control/robotics; despite arXiv venue and mixed h‑indices, author reputation warrants highest rating.