Heterogeneous robot teams can combine complementary sensing, mobility, and manipulation, but they often fail to learn from corrections made during execution, lose task-relevant context across space and time, and lack explicit representations of inter-robot dependencies. A new paper on arXiv introduces InstAgent, an agent operating system designed to address these three gaps.
InstAgent organizes team behavior as agentic behavior trees that explicitly assign capabilities to robots, model dependencies between them, and validate handoffs. A recursive self-improvement mechanism updates these collaboration programs from execution feedback and preserves successful revisions for reuse in future tasks. A shared multimodal spatio-temporal memory provides context for planning and handoff validation.
The authors tested InstAgent on 20 real-world multi-robot tasks using three high-level models, reporting consistent improvements in task success, first-attempt completion, and execution efficiency. The paper is available on arXiv under identifier 2610.23098.