Answers are grounded in this paper and cite their sources.
Polishing robots are widely used in industrial surface-finishing tasks, but collision detection remains challenging because labeled collision data are costly and potentially hazardous to collect. This paper proposes a self-supervised collision-detection framework that learns temporal representations from unlabeled robot interaction data without requiring collision labels during pre-training. A modality-specific dual-encoder architecture based on Gated Recurrent Units separately models joint motion and current-related signals, while intra-modal contrastive objectives and cross-modal alignment capture temporal consistency and complementary dynamic information. After pre-training, both encoders are frozen and only a lightweight probe classifier is trained for downstream collision detection. Task-oriented signal augmentations are further introduced to improve robustness to polishing-related disturbances. On a highly imbalanced simulation benchmark, the proposed method achieves an Area Under the Precision–Recall Curve (PR-AUC) of \(94.02%\) , a Best F1-score of \(93.88%\) , and a Recall of \(94.77%\) under a false-positive-rate constraint of 0.5% (R@FPR \(≤\) 0.5%), outperforming the compared methods across the four reported metrics. In controlled real-world proof-of-concept experiments on an RM65 robotic arm, few-shot probe adaptation improves window-level Recall from 73.44% to 86.08% and trial-level Recall from 60.00% to 79.20%. These results demonstrate the potential of self-supervised multimodal representation learning for collision detection under polishing-related conditions, while indicating that further validation is needed for practical deployment.
No takes yet. Share an insight, caveat, or question.
Xing et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: