Population Contracts for Scalable Language-Guided Multi-Robot Planning and Control

Large heterogeneous robot teams must translate natural-language missions into coordinated behavior without making planning and control representations depend on a fixed list of robot identities. Existing language-model planners often enumerate robots, while centralized multi-agent reinforcement learning commonly concatenates agent states; both choices complicate transfer to new team sizes. We propose a shared population-level interface between language planning, task allocation, and decentralized control. A Population Contract specifies group goals, roles, proportional quotas, capability requirements, priorities, constraints, and recovery conditions without naming robots. A validator and capability-aware optimizer ground the contract to available agents, after which a role-conditioned population policy maps local observations, assigned contract entries, capabilities, and group statistics to continuous actions. We establish conditional grounding soundness, identity-equivariance of the actor, identity-invariance of the critic, and population-independent network dimensions for a fixed schema. The implemented stack integrates contract validation, CP-SAT grounding, VMAS, BenchMARL training, and deterministic exact-population evaluation. In the current (N=10) coverage slice, MAPPO and the initial concatenation-based Group-MF both complete 500k-frame training, but the latter achieves only 0.12/0.05 static/dynamic success versus 0.21/0.09 for random control. All nine subsequent fixed-data actor runs pass a preregistered imitation-and-rollout gate, and a behavior-cloning warm start raises seed-0 gated-residual Group-MF success to 50/50 episodes. Thus, the present evidence validates the cross-layer implementation and localizes the immediate bottleneck to on-policy discovery; it does not yet establish language-cost scaling or unseen-population transfer.