Reliable Robotic Task Execution in the Face of Anomalies

TL;DR

The paper tackles the fragility of learned robotic policies in open environments by introducing a policy-agnostic framework that couples online visual anomaly detection with a three-stage recovery process. An anomaly detector trained on nominal executions uses self-supervised image features and nearest-neighbor distance to flag deviations, triggering recovery steps: pause ($ER_1$), local perturbation ($ER_2$), and restart from a learned successful-start distribution ($ER_3$) modeled as a Gaussian Mixture Model. The detector is evaluated with both dedicated and general-purpose policies, showing substantial boosts in execution success under anomalies, and the restart step often provides the strongest recovery signal. The results demonstrate practical robustness improvements for sim-to-real transfer and highlight the framework’s applicability across policy types, with open-source implementation and clear avenues for enhancement such as multimodal sensing and better termination criteria.

Abstract

Learned robot policies have consistently been shown to be versatile, but they typically have no built-in mechanism for handling the complexity of open environments, making them prone to execution failures; this implies that deploying policies without the ability to recognise and react to failures may lead to unreliable and unsafe robot behaviour. In this paper, we present a framework that couples a learned policy with a method to detect visual anomalies during policy deployment and to perform recovery behaviours when necessary, thereby aiming to prevent failures. Specifically, we train an anomaly detection model using data collected during nominal executions of a trained policy. This model is then integrated into the online policy execution process, so that deviations from the nominal execution can trigger a three-level sequential recovery process that consists of (i) pausing the execution temporarily, (ii) performing a local perturbation of the robot's state, and (iii) resetting the robot to a safe state by sampling from a learned execution success model. We verify our proposed method in two different scenarios: (i) a door handle reaching task with a Kinova Gen3 arm using a policy trained in simulation and transferred to the real robot, and (ii) an object placing task with a UFactory xArm 6 using a general-purpose policy model. Our results show that integrating policy execution with anomaly detection and recovery increases the execution success rate in environments with various anomalies, such as trajectory deviations and adversarial human interventions.

Figures (6)

• Figure 1: An overview of our proposed framework for failure-aware policy execution. During execution, visual anomalies are detected using a self-supervised feature extraction method and different recovery actions are attempted (pausing, perturbation, and finally a reset using information from a learned success model).
• Figure 2: Examples of anomalies in the case of a robot attempting to reach a door handle (wrist camera view)
• Figure 3: Experiment setup of the door handle reaching task in a simulated environment (left) and a real environment (right)
• Figure 4: Illustration of the results of the anomaly detection process during policy execution in the door handle reaching task. The anomaly here is a temporary hand obstruction as illustrated in Fig. \ref{['fig:anomaly-examples']} (successfully detected in this case).
• Figure 5: An example of a false negative detection in the door handle reaching task, where a collision of the end effector with the handle is not flagged as an anomaly.