How Many Bits Does a Swarm Need?
IROS 2026 Workshop on Insect-Scale Autonomy · Spotlight
Adaptive-Precision Task Allocation for Resource-Constrained Robots
Accepted for a spotlight presentation at the IROS 2026 Workshop on Insect-Scale Autonomy.
Motivation
Small robots operate with limited communication bandwidth, energy, and onboard computation. Task allocation often sends a 32-bit utility for every robot and task pair, even when much coarser values would lead to an equally valuable assignment. This project asks how much numerical precision a swarm actually needs to coordinate well.
Approach
Margin-Triggered Bit Refinement (MTBR) starts with coarse utility intervals. A coordinator solves the assignment and checks whether its worst-case value still exceeds that of the strongest competing assignment. If the bounds overlap, it requests extra bits only for utilities on which the assignments differ. Each bit narrows the uncertainty until the assignment can be certified or the bit budget is exhausted.
Four-bit utilities retain more than 98.8% of the 32-bit reference welfare with 87.5% less utility payload across the tested swarm sizes of 5 to 50 robots.
Welfare measures the total utility of the selected robot and task pairs. The exact pairing can change while the mission value remains nearly unchanged. MTBR also achieves lower welfare regret than random refinement and local-gap refinement at matched utility payloads.
Inside the spotlight slides
Selected visuals from the spotlight presentation. Download the complete editable 17-slide deck using the link above.
1 of 3. The question: How much communicated precision does a swarm need for good task allocation?
2 of 3. Value versus identity: Four-bit utilities preserve mission value even as reproducing the exact reference assignment becomes harder.
3 of 3. Where extra bits help: Targeting global assignment competition reduces welfare regret at matched utility payload.
What the results mean
The study evaluates synthetic spatial assignment problems, using 500 instances per swarm size for fixed precision and a separate 300 for adaptive refinement. The payload savings count utility bits. MTBR adds request and reply rounds, so lower payload does not automatically mean lower total network traffic. Hardware measurements and dynamic task allocation remain future work.
Comments