Safety Barrier Certificates for Heterogeneous Multi-Robot Systems

2 min read

Wang, Ames and Egerstedt add collision avoidance to any existing multi-robot controller as a minimally invasive filter. Each robot solves a quadratic program that stays as close as possible to its nominal input subject to pairwise control barrier function constraints, for double-integrator robots with bounded acceleration. In a mixed team, the responsibility for avoiding each collision is split in proportion to the robots’ acceleration limits, so agile robots do more of the swerving. When a robot does not know its neighbours’ limits, it starts from a conservative guess and refines it from local observations without giving up safety. The arXiv version extends the ACC paper with hardware experiments.

Findings

  • MATLAB simulation: six agents swap places across a circle. Small agents have 1.2 m/s² and a 0.2 m safety radius, the large one 0.6 m/s² and 0.4 m. The nominal goal-to-goal controller alone collides in the centre. With the certificates every agent reached its goal, the small ones did most of the avoiding, and the large one slowed but kept its course.
  • Hardware: three Khepera III robots (2.0 m/s²) and one Magellan Pro (0.5 m/s²), tracked by motion capture. The Khepera robots turned away while the Magellan pushed on, and all four reached the opposite side.
  • The controller is a QP per step with no reachable-set computation.

Limits

  • The authors admit deadlock: when objectives conflict, robots can stay safe and never finish. They also note that in crowded scenes several barrier constraints can make the QP infeasible, and guaranteed feasibility is open.
  • The distributed version assumes every neighbour runs the same filter and takes its share of the avoidance. A neighbour that ignores its share, through a bug or an attack, breaks the pairwise guarantee (my reading).
  • Small teams: six simulated and four real robots.
  • The journal follow-up, “Safety Barrier Certificates for Collisions-Free Multirobot Systems” (IEEE Transactions on Robotics, 2017), adds relaxed barriers, braking controllers and perturbations against deadlock. I have not read it.

Relevance

The standard multi-robot application of ames-2017-cbf-qp and the hand-written baseline that qin-2021-neural-barrier-certificates and bousias-2025-equivariant-cbf try to beat. The admitted deadlock is a safe-but-stuck state of the kind studied in Safe but Stuck. See safe control.