Learning Safe Multi-Agent Control with Decentralized Neural Barrier Certificates

3 min read

Qin and colleagues learn two neural networks per agent at once: a controller and a control barrier function (CBF) that certifies it. Each barrier depends only on the agent’s own state and what it observes of its neighbours, so the certificate is decentralised. If every agent satisfies its local barrier conditions, the paper proves (Proposition 1) that the whole team stays collision-free, with no central coordinator. The loss is a hinge penalty on the three CBF conditions (positive on safe initial states, negative in the dangerous set, non-decreasing fast enough elsewhere), evaluated on states sampled from closed-loop rollouts. Networks use a PointNet-style max-pooling over neighbours so they accept any number of neighbours in any order, and at test time an online policy refinement step nudges the action whenever the learned barrier’s derivative condition is about to fail.

Findings

  • Trained with 8 agents, tested with up to 1024 in 2D ground-robot tasks (Navigation, Predator-Prey, Nested-Rings) and 3D drone tasks (Maze, Tunnel), with agent density held roughly constant. The authors say 1024 was set by the compute of their laptop.
  • Safety rate with online refinement on Maze: 0.9999 (4 agents), 0.9987 (32), 0.9956 (1024). Without refinement: 0.9999, 0.9869, 0.9741. Tunnel follows the same pattern.
  • Refinement runs on under 17% of steps (0.1655 of steps at 1024 agents on Maze), and that fraction levels off as the team grows.
  • Beats MAMPS, PIC and MADDPG (plus safety-augmented variants) on safety rate and reward at 32 agents. The model-free baselines got 10^3 times more training samples.

Limits

  • The authors say plainly that it “still cannot guarantee that the agents are 100% safe”: the barrier conditions are enforced on training samples and can fail on unseen test states.
  • The generalisation bound is for the open-loop training distribution. Closed-loop distribution shift is left to future work, and the iterative retraining is only argued to make the gap negligible.
  • Dynamics are assumed known (the loss differentiates through ). Observations are assumed correct.
  • The certificate is a contract that every agent follows. One agent that ignores it, or lies about its state, breaks Proposition 1, and nothing in the method notices.

Relevance

MDBC is a learned, decentralised descendant of CBF-QPs and the hand-built multi-robot barriers of safety barrier certificates, and equivariant multi-agent CBFs start from it. For safe control in a team, the open question is what a barrier certificate is worth when a neighbour is Byzantine: the safety argument is only as good as each agent’s view of the others. See robust robotics.