Imagine you're tightening a jar lid in the dark. You can't see whether the lid is crooked, but the moment you start twisting, you feel resistance spike unevenly — harder on one side than the other. That asymmetry tells your wrist exactly which way to adjust. This paper does precisely that for a robot arm turning industrial valves, replacing cameras with torque sensors as the primary alignment signal. The committed claim: a provably stable haptic controller that corrects axial misalignment between a robot gripper and a valve handle using only the reaction torques generated during manipulation. The geometric insight is clean — misalignment between the gripper rotation axis and the valve axis produces off-diagonal reaction torques that are directly proportional to the error vector. The controller reads those torques and nudges the approach angle until they vanish. Uniform asymptotic stability is proven analytically, not just demonstrated. Why not just use a camera? Vision-based alignment is the dominant approach in robotic manipulation, but it carries a stack of failure modes: calibration drift, occlusion by the gripper itself, specular reflections on metal valves, and uncertainty about exact contact geometry. Torque feedback sidesteps all of these because it arises from the physical interaction itself — you're measuring the error at the point of contact, not inferring it from pixels meters away. The tradeoff is that you need to be already in contact to get a signal, which limits pre-contact planning. The architecture is a Lyapunov-based control law operating on SO(3) rotation geometry. The controller maps measured reaction torques through the known geometric relationship to produce corrective angular velocity commands. This is classical nonlinear control theory — no learning, no neural networks, no data-driven components. The stability proof is the payload: uniform asymptotic stability guarantees convergence regardless of initial conditions within the tested range, which is a stronger guarantee than most learning-based approaches can offer. Experimental validation uses a Kinova Gen3 7-DOF arm across 18 test cases: initial misalignments from −15° to +15° at three different valve positions. The headline result is a median residual misalignment of 2.46° and median residual reaction torque of 0.23 Nm. These are absolute-value medians across all cases, not cherry-picked best runs. That said, validation is entirely in-house — no independent replication, no community benchmark for valve manipulation, and the valve geometry appears to be a single standard type. The gap that matters: this is demonstrated on one valve type in a lab setting with a cooperative (non-stuck, non-corroded) valve. Industrial valve turning involves stuck valves requiring high torques, varying handle geometries, and environmental constraints like pipes and insulation blocking approach angles. The paper is honest about its scope — it's solving alignment correction, not the full manipulation pipeline — but the distance from lab demonstration to field deployment remains substantial. The next milestone to watch is whether this controller composes with force-controlled valve rotation (not just alignment) on valves that actually resist turning. The alignment problem and the torque-application problem interact: a stuck valve generates much larger reaction torques that could mask the alignment signal. If the method survives that coupling at industrial torque levels, it becomes genuinely useful for inspection robots in oil-and-gas or nuclear facilities.
