SKiD-SLAM: Resource-Aware Distributed Multi-Robot LiDAR SLAM for Planetary Missions with Field Deployments at Analog Sites

2 min read

A field report on distributed multi-robot LiDAR SLAM for planetary missions, where there is no central infrastructure, little onboard compute and very little bandwidth. The authors survey real mission budgets (Curiosity’s direct-to-Earth link runs at 500 bps to 32 kbps; Ingenuity talked to Perseverance over a 20 or 250 kbps Zigbee radio) and fix a budget of 250 kbps between robots within 1 km. They then pick a module for each job that fits: SOLiD descriptors for inter-robot place recognition, KISS-Matcher for registration, a truncated-MSE check and PCM for inter-robot outlier rejection, and distributed pose-graph optimisation. The system is checked on public datasets and a simulator built on JPL’s Open Source Rover, then deployed twice in South Korea: wheeled rovers in the Majang lava-tube cave on Jeju, and quadrupeds doing a crater-and-rock science scenario on analog terrain at KICT in Yeoncheon.

Findings

  • Distributed LiDAR SLAM ran reliably in both deployments within the 250 kbps inter-robot budget.
  • On Intel NUCs the full pipeline averaged 100 to 200 ms per step, with occasional peaks above 500 ms during inter-robot loop closure.
  • The paper ablates each module (descriptor, registration, truncated MSE, PCM, overlap between robots, communication robustness), and code and field data are released.

Limits

  • Each site used one kind of robot, so loop closure between wheeled and legged robots, whose LiDARs sit at different heights, is untested in the field (authors).
  • No formal robustness guarantee: PCM depends on its pairwise tests and the truncated-MSE threshold was chosen by hand (authors).
  • Earth analog sites differ from other planets in ways the paper discusses but cannot test.

Relevance

A rare collaborative-slam system designed from a bandwidth budget rather than tuned until it fits. Outlier rejection is PCM with no certificate, and the authors say TEASER++ was too heavy for the computers on board, which is the trade-off in robust-estimation. Nothing here handles a teammate that sends wrong data on purpose: PCM treats bad loop closures as noise. Compare SGM-SLAM for another way to save bandwidth.