At ICRA 2026 in Vienna, Daotong Technology's AvantRobotics team and the University of Macau unveiled the Implicit Virtual Leader framework, a decentralized vision-only formation control system.
ICRA 2026, under the theme Robots for All, gathered over 8,000 researchers from 86 countries with 4,947 submissions. Chinese institutions dominated, with Tsinghua University ranking first with 74 accepted papers.
The IVL framework addresses three critical challenges: single-point-of-failure vulnerability, GPS dependency, and lack of uncertainty awareness.
First Breakthrough: Elimination of leader-follower single-point failure. IVL introduces a non-physical formation centroid as reference, with no robot designated as leader. Each robot infers its 6DoF pose using only monocular images and inter-robot communication.
Second Breakthrough: Uncertainty awareness. The framework equips pose estimators with aleatoric uncertainty (via heteroscedastic Gaussian NLL loss) and epistemic uncertainty (via MC Dropout), achieving Expected Calibration Error of 0.022.
Third Breakthrough: Graph Neural Network architecture using two-stage Transformer design. Trained on 2-5 robot formations, it maintains stable performance on 6-7 robot formations with position error 0.29m and orientation error 9.3-10.3 degrees.
The framework has broad applications in inspection scenarios requiring multi-robot coordination in GPS-denied environments and heterogeneous air-ground robot teams.
