Bagla's self-calibration algorithm keeps robots on target with 100% success

In a paper submitted to arXiv (2608.12528), Yash Bagla tackles a tricky robotics problem: a vehicle must seek a hidden target using a range-bearing relay whose position and yaw are unknown. Prior work showed that certain vehicle-relative observations can remove the calibration gauge, but only in hindsight. Bagla supplies the missing closed-loop layer — a controller that decides online whether calibration is trustworthy enough to proceed with target seeking, or whether to trigger extra exploratory motion first.

At the core is a trajectory-spread margin, S_v, which simultaneously serves as a seed-accuracy bound, a variance decomposition, and an excitation budget. The supervision rule uses this margin to project away from exploratory pushes when calibration is sufficient, and otherwise retriggers exploration. In the noiseless regime, the method provably acquires required excitation in finite time and achieves target-seeking convergence after certification. The threshold is tied directly to a desired calibration-accuracy level, not chosen heuristically.

Validation includes closed-loop simulation, Monte Carlo comparisons, and a ROS 2/Gazebo software-in-the-loop experiment with sensing delay. The key result: a decay-rate sweep shows that when a fixed excitation schedule decays too fast, baseline yaw RMSE balloons from 0.010 to 0.065 rad and success collapses to 56%. The supervised approach keeps yaw RMSE between 0.0095 and 0.0191 rad and achieves 100% success — while target-tracking error stays insensitive. This makes the algorithm practical for drones, underwater vehicles, and other autonomous systems that must operate through uncertain relay nodes.