Robot swarms neutralize harmful Byzantine robots using a blockchain-based token economy

3 min read

Strobel, Pacheco and Dorigo run an Ethereum blockchain on up to 24 Pi-puck robots and use a smart contract as the swarm’s shared memory for a collective-sensing task: estimate the fraction of white tiles on a chequered floor. Every robot holds crypto tokens, and submitting an estimate costs tokens. The contract refunds and rewards estimates that pass its outlier filter, pays nothing for rejected ones, and hands every robot a “universal basic income” at exponentially growing intervals. A robot that keeps sending bad values runs out of tokens and can no longer influence the swarm. The chain uses proof-of-authority consensus (a majority of preselected robots must agree), because proof-of-work is too expensive for small robots.

Findings

  • With no Byzantine robots the median absolute error was 0.6 % on real robots and 0.6 % in simulation, against 0.5 % for an offline baseline, so the security layer costs almost nothing when nobody lies.
  • With 9 Byzantine robots out of 24 (37.5 %) always reporting “0 % white”, the median error stayed below 2.2 %. The baseline error grew roughly linearly with the number of Byzantine robots.
  • More Byzantine robots meant slower convergence, because rejected estimates leave fewer valid ones per unit time.
  • Validated simulations scaled to 120 robots and 600-minute runs. The chain grew linearly: 73.6 MB after 600 minutes with 120 robots, mean bandwidth 2.11 KB/s, under 1 % of the Pi-puck’s Wi-Fi capacity.
  • Every robot, Byzantine ones included, ends up holding the same tamper-evident log, so one surviving robot is enough to reconstruct a run.

Limits

  • The authors say their contract is simple and could be circumvented by smarter Byzantine robots. Detection is an outlier test on one scalar in a static environment.
  • The swarm is closed and its size known in advance. Once robots can join or leave, Sybil attacks become possible, and a permissioned chain needs someone to grant permissions (authors).
  • Network partitions could let an attacker hold a relative majority inside a sub-swarm (authors). Proof-of-authority itself needs an honest majority of sealers.
  • The paper defines a Byzantine robot as one whose behaviour differs from its intended behaviour, whether from bugs, failed parts or attacks. The filter treats a broken ground sensor and a liar the same way, and the only response is economic exclusion.
  • Communication range was 10 cm, and the authors admit a single robot could do the task.

Relevance

It takes strobel-2020-blockchain-consensus-swarms onto physical robots. It swaps the per-message trimming of W-MSR for a budget: a liar may speak until it has spent its tokens, which suits slow, high-stakes swarm decisions more than fast control loops. See Byzantine faults and robot security.