Autonomous robotic refuelling
Centre for Systems and Control, IIT Bombay · Jan 2026 – Present
Bachelor thesis
Abstract
Refuelling is an insertion task with no room for a retry. The nozzle must reach a port whose position is known only approximately, in a workspace with obstacles, and make contact without loading the structure it is entering. The architecture avoids both iterative inverse-kinematics solvers and heavy sampling-based global planners, transitioning between configuration-space and Cartesian tracking within a single pipeline.
The problem
An insertion task under uncertainty punishes a controller that treats position error as the only thing worth driving to zero. Push hard enough on a misaligned port and the arm succeeds at its tracking objective by damaging the thing it was inserting into.
Iterative inverse kinematics gives no bound on solve time and no guarantee about which solution it lands on. A sampling-based global planner is too heavy to sit in a reactive loop. The architecture is built to avoid depending on either.
The stack
Algebraic inversion via Paden-Kahan subproblems gives exact, constant-time spatial tracking rather than iterating toward a solution. A derivative-based selector then chooses among the valid solutions by evaluating velocity, acceleration and jerk over a rolling history buffer, so the arm does not jump between branches of the inverse map mid-motion.
STOMP, a gradient-free stochastic optimizer, generates the smooth collision-free gross motion. Bubble strips handle what appears after planning, deflecting the path in real time through Jacobian-transpose mapping instead of replanning from scratch.
The final approach is governed by an admittance layer, a mass-damper-stiffness law driven by external force, which maps measured force to joint motion so the nozzle yields to the geometry it meets rather than forcing through it.
Validation
The stack runs as a four-phase mission orchestrator on ROS 2 Jazzy with Gazebo, and was exercised in simulation against randomized obstacle fields before going near hardware. Hardware validation is on a Kinova 6-DOF arm. This became my bachelor thesis.