CVPR 2026 RC NF

RC-NF — Robot-Conditioned Normalizing Flow for Real-Time Anomaly Detection in Robotic Manipulation

Venue: CVPR 2026 Category: Anomaly Detection / Robustness Trend tag: Trend 6 Affiliations: Fudan ITEA + SMU Authors: Shijie Zhou, Bin Zhu, Jiarui Yang, Xiangyu Zhao, Jingjing Chen, Yu-Gang Jiang

Approach diagram

flowchart LR
  OBS["observation"] --> SAM2["SAM2 mask<br/>→ point sets"]
  SAM2 --> P1["dynamic-shape branch"]
  SAM2 --> P2["positional-residual branch"]
  P1 --> NF["robot-conditioned<br/>coupling flow (RCPQNet)"]
  P2 --> NF
  ROBOT["robot state<br/>(joints, gripper, pose)"] --> NF
  TASK["task embedding (FiLM)"] --> NF
  NF --> SCORE["anomaly score (neg log-density)"]
  SCORE --> DEC["anomaly? &lt;100 ms"]

Loading

Problem

Anomaly detection on top of VLA policies has been slow and policy-specific. A general, real-time detector that can run on top of any VLA (π0, OpenVLA, RDT-class) is missing.

Method

A task-aware, robot-conditioned coupling normalizing flow (Glow-based, K=12 flow steps) with dual-branch point features:

• Use SAM2 to segment objects in the observation, then grid-sample masks into point sets. Two complementary branches encode them: a dynamic-shape branch that normalizes each frame to remove translation/scale, and a positional-residual branch that restores the absolute-position information lost in normalization.
• The core block is the Robot-Conditioned Point Query Network (RCPQNet), an affine coupling layer that decouples robot-state and object features while preserving their interaction: robot-state features (joint angles, gripper state, Cartesian pose) act as query tokens and object point features as memory tokens in cross-attention, with transformation parameters modulated by the task embedding via FiLM. Conditioning is on the robot's state, not robot identity — so anomaly scoring is grounded in whether robot and object motion stay consistent with the task.
• Trained unsupervised on positive (normal) samples only. Score new observations by their negative log-density under the flow; low density → anomaly, enabling state-level rollback or task-level replanning.

Results

Evaluated on LIBERO-Anomaly-10, a new simulation benchmark introduced by the paper covering three manipulation-specific anomaly types: Gripper Open (gripper stays open while grasping), Gripper Slippage (zero friction → object slips), and Spatial Misalignment (robot moves toward the wrong compartment).

• SOTA: 0.9309 average AUC / 0.9494 average AP, vs. FailDetect at 0.7181 AUC / 0.7700 AP. Baselines also include VLM-based detectors (GPT-5, Gemini 2.5 Pro, Claude 4.5).
• Sub-100 ms detection latency on a consumer GPU (RTX 3090), running as a plug-and-play monitor atop existing π0 / OpenVLA-class policies.

Significance

The first general, real-time, plug-in anomaly detector for VLAs. Critical infrastructure for safe deployment, especially as long-horizon VLAs proliferate. Sister role to SafeVLA and Latent Policy Barrier — both prevent unsafe actions; RC-NF detects unsafe states.

Links

• arXiv: 2603.11106
• Project: heikaishuizz.github.io/RC-NF

Related pages

• SafeVLA · Latent Policy Barrier · CycleVLA
• CVPR 2026 survey

← Back to CVPR-2026