← Research

MULTI-ROBOT SYSTEMS · 2026

AC-DC: Adaptive Communication for Scalable Dynamic Average Consensus

Adaptive peer-to-peer communication for multi-robot ergodic search under finite range, finite rate, and receiver-side interference.

ICRA 2027 submission · arXiv preprint 2026 Lead author

1 · Where should robots search?

Search target: where information matters

↓

Team visitation: where robots actually search

More important regions should receive more team search effort.

2 · AC-DC communication layer

Robots choose what, when, and with whom

Where I
have searched
What I
have sensed
Last information from nearby robots
AC-DC
WHATWhich information is most useful?
WHENIs it worth communicating now?
WHOWhich neighbor is best to contact?
Output: better shared estimates of team coverage and sensing information

3 · What improves?

Lower uncertainty with less modeled traffic

29.5%lower paired AUC vs ADMM-DAC
24.5%lower paired AUC vs PP-ACDC
5.8×ADMM-DAC / AC-DC traffic
11.5×PP-ACDC / AC-DC traffic
120 robots: 19.3 MB vs 19.2 MB
AC-DC vs ideal centralized reference

Why communication is part of the search problem

We study robot teams that gather information while they move. The mission priority specifies where reducing uncertainty matters most, while team visitation describes where the robots actually search. Coordinated ergodic search therefore needs shared information about both where the team has searched and what the team has sensed.

The network is not an unlimited background service: communication has finite range and data rate, and nearby transmissions can interfere. We keep the receding-horizon ergodic controller fixed and instead adapt the communication policy that supplies its distributed information.

What AC-DC adapts

Each robot tracks two moving team averages: trajectory statistics for team visitation and regional sensing information used to update the search target. AC-DC uses local inputs, successfully received neighbor information, and stored neighbor blocks to decide:

What

Select the state block whose exchange is predicted to be most useful.

When

Request communication when the exchange score is high enough, with an age-based refresh fallback.

Who

Among qualified neighbors, request the neighbor with the largest score for the selected block.

The exchange decision is made before receiving the neighbor’s current block. Robots use stored neighbor blocks and compact drift summaries as a scoring proxy. When both directions of an ordinary exchange succeed, the pair averages the exchanged block and refreshes its neighbor memory.

Dynamic-priority search experiment

The mission priority changes at 100 s, so the regions that matter most also change during the run. We evaluate closed-loop search using the normalized priority-weighted remaining covariance-trace ratio and summarize each 250 s trajectory by its time-averaged AUC. Lower AUC means less remaining mission-relevant uncertainty.

12map-size / density settings
20paired trials per setting
80robots in the main study
120robots in fixed-area scaling

Results

Across all twelve settings, AC-DC had the lowest mean AUC and attempted modeled payload among the compared decentralized methods. The paired aggregate reductions were 29.5% versus ADMM-DAC and 24.5% versus PP-ACDC. Those baselines used 5.8× and 11.5× AC-DC’s modeled traffic, respectively.

In the fixed-area scaling study, selected-block AC-DC used 19.3 MB at 120 robots versus 19.2 MB for the ideal centralized reference. This is a communication-scaling comparison, not a claim of centralized-level search performance: the corresponding AUCs were 0.129 and 0.041.

Takeaway

AC-DC treats communication as a decision inside the autonomy stack. Rather than asking every robot to exchange every state whenever possible, each robot adapts What, When, and Who using information it can maintain locally.