Self-governed access control in open robot swarms using a blockchain-based DAO

3 min read

Earlier blockchain swarms assumed fixed membership. This paper looks at open swarms, where robots join and leave at runtime, and lets the swarm govern who is in it through a decentralised autonomous organisation (DAO): rules in smart contracts on a blockchain shared by the robots. The proof of concept regulates membership. Robots join and must give valid estimates for a while before they earn the right to vote on removals. Every so often each voting robot picks removal candidates with its own private rule, and the smart contract only counts votes and enforces the majority. Keeping each robot’s detection rule private means a Byzantine robot cannot read a public threshold and stay just under it. The task is the usual collective sensing one from the token-economy paper: estimate the share of white floor tiles (25%), here with the tiles clustered in two corners to make it harder.

Findings

  • Motivation for open swarms: in a larger arena, a static swarm of 12 robots struggles to estimate the tile share, while adding honest robots every 120 s up to 30 robots drives the error towards zero.
  • Honest robots flag suspected collusion (three or more robots sending identical or near-identical estimates) first, and otherwise outliers by an adaptive threshold. Smart Byzantine robots invert this and target the honest robots closest to the mean.
  • When every removal triggers two new Byzantine robots, a longer waiting period before newcomers can vote gives more true positives and fewer false positives. With a short wait, newly added Byzantine robots can nominate honest robots almost at once.
  • The simulations (ARGoS with the Toychain blockchain, simulated Pi-pucks) keep the swarm working with the number of Byzantine robots within acceptable limits. The abstract gives no single headline number.
  • Real robots were used only for the hardest case, preventing a Byzantine takeover of a healthy swarm, with the setup sped up (detection every 150 time-steps, about 2.5 minutes) to fit battery life and the number of robots available.

Limits

  • The swarm survives a Byzantine majority only if the Byzantine robots have not yet earned voting rights and the honest ones expel them first (authors).
  • The detection rule is tuned to one scalar estimation task. How it would carry to maps or plans is not tested.
  • Most results are in simulation, with real robots as a demonstration of one case.

Relevance

This continues blockchain consensus and the token economy from the same group, and moves the question from “who is lying” to “who is allowed to vote”. The waiting period before newcomers vote is a cheap Sybil defence, and it is worth comparing with the physical checks in Sybil4Drones and the per-link tests in the Safe Wrapper. See byzantine-fault-tolerance and comparison-byzantine-defences.