Industrial deployment of robotic visual anomaly detection (VAD) is fundamentally constrained by passive perception under diverse 6-DoF pose configurations and unstable operating conditions such as illumination changes and shadows, where intrinsic semantic anomalies and physical disturbances coexist and interact. To overcome these limitations, a paradigm shift from passive feature learning to Active Canonicalization is proposed. PiCo (Pose-in-Condition Canonicalization) is introduced as a unified framework that actively projects observations onto a condition-invariant canonical manifold. PiCo operates through a cascaded mechanism. The first stage, Active Physical Canonicalization, enables a robotic agent to reorient objects in order to reduce geometric uncertainty at its source. The second stage, Neural Latent Canonicalization, adopts a three-stage denoising hierarchy consisting of photometric processing at the input level, latent refinement at the feature level, and contextual reasoning at the semantic level, progressively eliminating nuisance factors across representational scales. Extensive evaluations on the large-scale M2AD benchmark demonstrate the superiority of this paradigm. PiCo achieves a state-of-the-art 93.7% O-AUROC, representing a 3.7% improvement over prior methods in static settings, and attains 98.5% accuracy in active closed-loop scenarios. These results demonstrate that active manifold canonicalization is critical for robust embodied perception.
Figures & tables
Figure 1: Passive Learning vs Active Canonicalization: PiCo Framework for VAD. Conventional visual anomaly detection relies on passive feature learning under fixed viewpoints and illumination, resulting in limited robustness to pose variations, uneven lighting, and specular artifacts. In contrast, PiCo integrates Active Physical Canonicalization and Neural Latent Canonicalization to actively reduce condition-induced variability, achieving 98.5% accuracy under closed-loop operation.
Figure 2: Multi-stage Canonicalization Pipeline of the PiCo Framework. This cascaded architecture progressively disentangles semantics from physical nuisances across three neural levels: photometric (Stage I), latent spectral (Stage II), and global contextual (Stage III). The resulting invariant representations ultimately drive the active physical canonicalization policy.
Figure 3: Modular Architecture of PiCo’s Canonicalization Mechanism. The framework integrates a photometric preprocessing module for lighting estimation, a dual-path bottleneck MLP for latent feature refinement, and contextual canonicalization. Ultimately, reconstruction uncertainty serves as a real-time feedback signal, driving the robotic agent to actively seek a geometrically optimal canonical view.
Figure 4: Representative cases from the geometrically complex subset of PiCo-Bench. Each block compares normal and anomalous samples, with subtle structural defects marked by red dotted circles. Their strong non-planarity induces pose-dependent & self-occlusion, exemplifying semantic–nuisance entanglement.
Category
Dinomaly [ 16 ] CVPR’2025
RD++ [ 12 ] CVPR’2023
MSFlow [ 13 ] TNNLS’2024
INP-Former [ 17 ] CVPR’2025
PiCo (Ours)
Improvement (vs Dino)
Bird
75.1/74.9/86.9
90.3 /70.2/79.8
85.0/62.0/71.4
80.0/67.2/84.1
80.7/76.7/90.9
+5.6 / +1.8 / +4.0
Car
86.7/75.1/78.3
85.0/68.2/75.6
67.9/55.9/67.4
58.1/53.9/72.1
89.3/84.6/93.0
+2.6 / +9.5 / +14.7
Cube
82.3/77.8/86.0
83.1/74.6/80.7
66.0/57.8/58.7
77.9/74.5/80.6
96.8/82.4/92.6
+14.5 / +4.6 / +6.6
Dice
98.1/93.0/85.7
98.4 /89.4/85.6
76.8/69.4/77.0
93.3/83.7/87.7
96.7/94.0/90.5
-1.4 / +1.0 / +4.8
Doll
74.4/72.6/89.0
66.8/65.9/85.4
56.4/55.1/68.9
72.5/73.7/85.8
81.6/78.6/90.9
+7.2 / +6.0 / +1.9
Motor
95.4/85.4/94.2
92.2/87.9/94.9
86.0/61.4/86.7
83.7/61.1/91.9
99.6/92.2/95.5
+4.2 / +6.8 / +1.3
Table 1: Quantitative comparison on M 2 AD. PiCo significantly outperforms the previous SOTA (Dinomaly) by +3.7%, +2.7% and +6.5% on average.
Figure 5: Qualitative anomaly detection results on the M2AD benchmark. The rows display normal image, anomalous image, PiCo’s predicted anomaly map masked by ground truth. Evaluated under highly fluctuating illumination and viewpoints, PiCo precisely localizes semantic defects while suppressing environmental nuisances.
Variant
bMlp
DIR ( α=0.1 )
DIR ( α=0.2 )
LA3
AUROC-O
AUROC-I
AUPRO
(a) (Baseline)
✓
93.33
86.51
91.77
(b)
✓
✓
93.35
86.46
92.06
(c)
✓
✓
93.00
86.19
91.97
(d) (Full)
✓
✓
✓
93.72
85.68
92.40
Table 2: Ablation Study on Neural Components. Dual-path bMlp with DIR( α=0.1 ) + Light-Aware Linear Attention shows optimal balance for manifold constraints.
Figure 6: Quantitative evaluation of active canonicalization on 50 hard cases. (a) Uncertainty collapse: Active robotic re-orientation forces widely dispersed, high-uncertainty observations to converge into a tight, low-uncertainty manifold. (b) Accuracy comparison: While passive SOTA baselines fail under severe physical occlusions (peaking at 68.2%), PiCo’s active pipeline breaks this physical ceiling, achieving 98.5% accuracy at the canonical view.
Recent advances in Vision-Language-Action (VLA) models have enabled robots to execute increasingly complex tasks. However, VLA models trained through imitation learning struggle to operate reliably in dynamic environments and often fail under Out-of-Distribution (OOD) conditions. To address this issue, we propose Robot-Conditioned Normalizing Flow (RC-NF), a real-time monitoring model for robotic anomaly detection and intervention that ensures the robot's state and the object's motion trajectory align with the task. RC-NF decouples the processing of task-aware robot and object states within the normalizing flow. It requires only positive samples for unsupervised training and calculates accurate robotic anomaly scores during inference through the probability density function. We further present LIBERO-Anomaly-10, a benchmark comprising three categories of robotic anomalies for simulation evaluation. RC-NF achieves state-of-the-art performance across all anomaly types compared to previous methods in monitoring robotic tasks. Real-world experiments demonstrate that RC-NF operates as a plug-and-play module for VLA models (e.g., pi0), providing a real-time OOD signal that enables state-level rollback or task-level replanning when necessary, with a response latency under 100 ms. These results demonstrate that RC-NF noticeably enhances the robustness and adaptability of VLA-based robotic systems in dynamic environments.
Shijie Zhou, Bin Zhu, Jiarui Yang +3
Institute of Trustworthy Embodied AI, Fudan University · Shanghai Key Laboratory of Multimodal Embodied AI · Singapore Management University
Physical active vision allows robots to change their viewpoint when task-relevant observations become unreliable, yet existing manipulation benchmarks provide limited support for studying how policies recover from occlusion during execution. We introduce BAVO-Bench (Bimanual Active Vision under Occlusion), a bimanual active-vision benchmark that systematically controls external visibility through Clean, Stage Occlusion, and Random-time Occlusion conditions, enabling evaluation of both manipulation performance and active visual recovery. Building on this setting, we present A-FAR (Active Future-Aware Recovery), an active-vision policy for joint viewpoint and manipulation control. A-FAR represents moving-camera observations in a unified robot-centric 3D frame and distills relational structure together with its future evolution from a pretrained 4D model, providing the policy with future-aware geometric guidance without requiring future observations at deployment. Experiments across multiple manipulation tasks show that A-FAR improves robustness to both structured and temporally shifted occlusions while maintaining strong performance under clean observations.
Kaijun Luo, Yudi Huang, Qijun Zhong +3
Sun Yat-sen University · University of Electronic Science and Technology of China · X-Era AI Lab +1
Video anomaly detection (VAD) is an actively studied task, having wide applications in typical scenarios such as public surveillance and road traffic safety. The task is also relevant for robotic arm interactions, where it has several downstream applications, including learning better interaction and manipulation abilities, triggering recovery procedures when anomalies occur, etc. Despite its relevance, the exploration of anomaly detection in robotic arm manipulation videos is limited by the low number of available resources. To this end, we introduce RoboVAD, a large-scale benchmark for video anomaly detection that comprises challenging cross-domain evaluation scenarios, where certain actions (tasks executed by a robotic arm) and anomaly types (mistakes that occur while performing certain tasks) remain unseen during training. RoboVAD is designed to benchmark VAD methods in realistic scenarios, where robotic arms can perform unforeseen tasks, and thereby encounter new anomaly types. We train and evaluate several state-of-the-art VAD methods, including a novel method specifically adapted for robotic arm manipulation. While the proposed method outperforms many state-of-the-art competitors, all methods remain below a micro-averaged frame-level AUC threshold of 70% in the most challenging evaluation setup, confirming the difficulty of the proposed benchmark. We publicly release our dataset and code at https://zenodo.org/records/22754659.
Alexandru-Bogdan Dura, Sebastian Balmus, Radu Tudor Ionescu
ICI Bucharest, Romania · University of Bucharest, Romania