Robots Master Curved Surfaces with New Two-Stage Method

Robots performing tasks like polishing, coating, or inspection on curved surfaces often struggle with maintaining consistent contact and smooth motion, leading to errors and inefficiencies. A new two-stage framework developed by researchers at the University of Tennessee, Knoxville, addresses this by separating geometric adaptation from real-time execution, allowing robots to follow periodic motion patterns on non-planar surfaces with improved stability and reduced drift. This approach is particularly relevant for industries relying on surface finishing and guided interactions, where precision and repeatability are critical.

The key finding is that directly applying pre-designed motion primitives to curved surfaces, a common practice, causes geometric inconsistencies such as interpenetration, orientation discontinuities, and cumulative drift over repeated cycles. The researchers discovered that by first warping these motion patterns offline to match the surface geometry and then using online projection with contact feedback, robots can achieve stable, drift-free execution. For example, on concave surfaces, direct tiling led to collision-prone attitudes, while the warped avoided these failures, as shown in Figure 5 of the paper.

Ology involves two main operators: an offline surface-constrained warping operator and an online contact-aware projection operator. The offline warping takes a nominal periodic motion primitive and embeds it onto a curved surface through a four-step process: extracting and tiling waypoints along a guide curve, constructing a dual-track tool-axis parameterization, applying an asymmetric diffeomorphic deformation based on local surface geometry, and building a smooth pose sequence with axis-consistent orientation completion. This produces a surface-adapted reference trajectory, as illustrated in Figure 3. The online projection then uses force-sensing resistor (FSR) feedback to adjust the robot's pose along gravity and enforces a conic safety constraint to bound orientation deviation, ensuring persistent contact without drift, detailed in Figure 4.

From experiments across five analytic surface families (sin, cos, exp, parabolic, cubic) and real-robot validation on a sinusoidal surface demonstrate significant improvements. Offline, the warping reduced large rotational jumps, with the 95th-percentile angular step decreasing by up to 16.59 degrees on sinusoidal surfaces compared to direct tiling, and eliminated bad-step rates (steps with angular changes over 10 degrees) in most cases, as summarized in Table I. Online, the projection improved contact consistency under disturbances and recovered from abrupt height reductions, maintaining bounded orientation deviation, shown in Figure 7. The real-robot tests confirmed that the framework enhances robustness with minimal sensing, using only a 1D FSR signal.

Of this research are substantial for real-world applications in manufacturing, healthcare, and maintenance, where robots must perform repetitive tasks on varied curved surfaces. By decoupling geometric embedding from execution, enables safer and more reliable robotic operations without requiring extensive recalibration or sensor arrays. This could lead to more efficient surface finishing processes, better inspection accuracy, and enhanced surgical assistance tools, as the paper notes applications in polishing, coating, inspection, and surgical assistance. The framework's ability to handle diverse surface geometries, from convex to concave, makes it versatile for industrial and medical settings.

Limitations of the study include its focus on smooth surfaces and periodic motion primitives, which may not generalize to highly irregular geometries or non-repetitive tasks. The paper does not address scenarios with dynamic surfaces or multiple contact points, and the online projection relies on gravity direction and FSR feedback, which could be less effective in zero-gravity environments or with different sensor types. Additionally, the validation was conducted on a limited set of analytic surfaces and one real-robot setup, suggesting further testing is needed for broader applicability. The researchers acknowledge that while the framework improves continuity and safety, it does not fully eliminate all geometric s in complex, real-world environments.