⚡ electrical engineering

Cost-constrained selective robotic re-polishing of complex freeform surfaces driven by local quality risk evaluation

This paper proposes a cost-constrained selective robotic re-polishing method for complex freeform surfaces that utilizes a multimodal risk evaluation framework to identify and target only high-risk defective regions, thereby significantly improving surface pass rates while reducing re-polishing area, time, and path length compared to conventional full-area approaches.

Original authors: Miao Yu, xu liu, Baowen He, Zhen Pan

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Miao Yu, xu liu, Baowen He, Zhen Pan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master sculptor trying to polish a giant, bumpy, free-form statue made of steel. Your goal is to make the entire surface perfectly smooth, like a mirror.

In the old way of doing things, you would use a robotic arm to polish the entire statue from top to bottom. But here's the problem: because the statue has curves, bumps, and edges, the robot doesn't polish every spot equally. Some areas get super smooth, while others remain rough or "bumpy" because the robot's tool bounced or pressed too hard in those specific spots.

If you wanted to fix the rough spots, the traditional method was to just polish the whole statue again. This works, but it's a waste of time, money, and tool wear. You are spending hours polishing areas that were already perfect, just to fix a few tiny bad spots.

This paper proposes a smarter, "targeted" approach. Instead of polishing the whole statue again, the robot acts like a detective to find exactly which spots need fixing, and only polishes those.

Here is how they did it, broken down into simple steps:

1. The "Detective" System (Risk Evaluation)

The researchers built a system that acts like a high-tech detective. They gave the robot a "multisensor" toolkit:

Eyes: A 3D camera to see the shape of the statue.
Touch: Sensors to feel how hard the tool is pressing.
Motion Tracking: A system to watch exactly how the robot arm is moving.

As the robot polishes, this system breaks the process into tiny, overlapping "snapshots" (like frames in a movie). For every snapshot, it asks: "Based on the shape of this spot and how the robot is moving right now, is this area likely to end up rough?"

2. The "Weather Forecast" for Roughness

The robot doesn't just guess; it uses a special math model (called GB-DRC) to predict the future quality of the surface.

Think of the shape of the statue as the "terrain" (like a steep hill vs. a flat plain).
Think of the robot's movement as the "weather" (like wind speed or rain).

The model realizes that a steep hill (complex geometry) needs different handling than a flat plain. It separates the "terrain" problems from the "weather" problems to predict exactly how rough a specific spot will be after the first polish.

3. The "Top-K" List (Choosing Who to Fix)

Once the robot finishes the first polish, it doesn't look at the whole statue. Instead, it looks at its "forecast" and creates a Top-K list.

It ranks every small section of the statue by how "risky" it is (i.e., how likely it is to be too rough).
It picks the top 25 most risky spots out of 42 total spots.
It ignores the 17 spots that are already smooth enough.

4. The Result: Saving Time and Money

The team tested this on a real steel workpiece. Here is what happened:

The Old Way (Polishing Everything): If they polished the whole thing again, they would fix the rough spots, but they wasted 100% of their extra time polishing good spots.
The New Way (Selective Polishing): By only polishing the 25 risky spots (which covered about 60% of the total area), they achieved a 81% success rate (where spots met the quality standard).
- This was a huge improvement over the first pass (which was only 64% successful).
- Most importantly, they saved about 37-40% of the time, tool path length, and extra work compared to polishing the whole thing again.

The Bottom Line

This paper is about teaching robots to be efficient. Instead of blindly re-doing work that is already good, the robot uses data to identify exactly where the "bad" spots are and fixes only those. It's the difference from a gardener who waters the entire lawn just because one patch is dry, versus a gardener who uses a hose to water only the dry patch.

The result is a smoother surface with significantly less wasted effort and cost.

Technical Summary: Cost-constrained Selective Robotic Re-polishing of Complex Freeform Surfaces

Problem Statement
In the robotic polishing of complex freeform surfaces, nonuniform interactions between local geometry and process dynamics frequently result in uneven surface quality and residual defective regions after an initial polishing pass. While conventional industrial practice addresses this through full-area uniform re-polishing, this approach is inefficient and costly. It introduces substantial redundant processing time, excessive tool path lengths, and unnecessary tool wear, particularly in regions that have already met quality requirements. Existing studies on surface quality prediction typically focus on global estimation or endpoint evaluation, failing to provide the localized, actionable data necessary for targeted corrective polishing decisions. Consequently, a gap exists between predictive quality estimation and cost-constrained selective re-polishing planning.

Methodology
The paper proposes a risk-guided selective robotic re-polishing framework designed to balance surface quality enhancement with processing cost reduction. The methodology consists of three primary stages:

Multimodal Data Construction and Alignment:
- A multi-sensor robotic platform was established to acquire initial point cloud geometry, real-time force/torque signals, Tool Center Point (TCP) motion states, and offline roughness measurements.
- The continuous polishing process was segmented into sliding time windows (0.4s length, 0.2s step) and aligned spatially to a unified workpiece coordinate system.
- A regional attribution strategy was employed to map window-level process states to specific local regions (42 regions total) based on trajectory coverage ratios, creating a dataset of 5,521 valid window samples.
Geometry-Guided Quality Risk Modeling (GB-DRC):
- A Geometry-guided Residual Correction (GB-DRC) model was developed to predict local roughness changes ( $\Delta Ra$ ).
- The model decomposes roughness evolution into two components: a geometric base trend (driven by curvature, boundary distance, normal variation, and initial roughness) and a process dynamic residual (driven by contact force fluctuations, TCP motion variations, and process parameters).
- An adaptive weighting mechanism regulates the contribution of the process disturbance term based on local operating conditions, enhancing prediction robustness across varying geometries.
- The model was trained using a grouped isolation strategy to prevent data leakage, ensuring evaluation under variations in batches, regions, difficult geometries, and process conditions.
Risk-Coverage-Oriented Selection Strategy:
- Window-level predictions are aggregated to the regional scale using coverage-ratio weighting to estimate post-processing roughness.
- A regional quality risk score is calculated based on the deviation from a target roughness threshold ( $Ra_{target} = 0.250 \mu m$ ).
- A Top-K selection strategy prioritizes candidate regions for re-polishing based on these risk scores. The strategy prioritizes "risk coverage" (capturing actual defective regions) over absolute precision to ensure no substandard areas are missed, accepting a conservative inclusion of some compliant regions.

Key Contributions

Selective Re-polishing Framework: The paper introduces a framework that enables targeted corrective polishing for complex freeform surfaces, moving away from inefficient full-area uniform reprocessing.
Multimodal Local Quality Evaluation: It develops a robust evaluation method integrating geometric characteristics, in-process dynamic responses, and TCP motion information to estimate localized roughness variations under varying conditions.
Actionable Decision Strategy: It establishes a risk-coverage-oriented candidate region selection strategy that converts localized quality evaluations into actionable corrective plans, effectively trading off quality enhancement against processing cost.

Experimental Results
Experiments were conducted on a Q235 steel workpiece with a complex freeform surface using a multi-sensor robotic platform. The results were validated across five grouped isolation testing scenarios (batch-independent, region-group independent, difficult-geometry, process-condition independent, and region-condition extrapolation).

Prediction Performance: The GB-DRC model demonstrated superior stability and lower prediction errors (MAE, RMSE) compared to baseline models (RF, XGBoost, GRU) across all testing scenarios, particularly in difficult-geometry and process-condition variations.
Region Selection: With $K=25$ candidate regions selected (covering 59.52% of the total area), the method achieved a risk coverage of 93.33% (identifying 14 out of 15 actual defective regions) and a candidate precision of 56.00%.
Quality and Cost Trade-off:
- Surface Quality: The selective re-polishing strategy (S3) improved the surface pass rate from 64.29% (one-pass polishing) to 80.95%.
- Cost Reduction: Compared to full-area uniform re-polishing (S2), the proposed selective strategy reduced the re-polishing area by 40.48%, additional processing time by 37.67%, and additional path length by 37.44%.
- The final mean roughness for the selective strategy was $0.220 \pm 0.002 \mu m$ , significantly better than one-pass polishing ( $0.258 \mu m$ ) while avoiding the full cost of uniform re-polishing.

Significance and Claims
The paper claims that the proposed method provides a practical and cost-effective solution for localized quality enhancement in robotic polishing. By shifting from global quality estimation to localized risk evaluation, the framework addresses the inefficiency of full-area reprocessing. The core value lies in the ability to transform window-level process predictions into actionable, selective re-polishing plans. The authors assert that this approach achieves an effective trade-off between surface quality improvement and processing cost reduction, making it suitable for cost-sensitive robotic finishing applications. The framework is noted to be adaptable to other freeform finishing tasks, provided that model parameters are re-calibrated for specific materials and tool conditions.