ARMADA: Autonomous Online Failure Detection and Human Shared Control Empower Scalable Real-world Deployment and Adaptation

Wenye Yu; Jun Lv; Zixi Ying; Yang Jin; Chuan Wen; Cewu Lu

cs.ROOct 2, 2025

ARMADA: Autonomous Online Failure Detection and Human Shared Control Empower Scalable Real-world Deployment and Adaptation

Authors: Wenye Yu, Jun Lv, Zixi Ying, Yang Jin, Chuan Wen, Cewu Lu

Organizations: Shanghai Jiao Tong University · Shanghai Innovation Institute · Noematrix Ltd.

Abstract

Imitation learning has shown promise in learning from large-scale real-world datasets. However, pretrained policies usually perform poorly without sufficient in-domain data. Besides, human-collected demonstrations entail substantial labour and tend to encompass mixed-quality data and redundant information. As a workaround, human-in-the-loop systems gather domain-specific data for policy post-training, and exploit closed-loop policy feedback to offer informative guidance, but usually require full-time human surveillance during policy rollout. In this work, we devise ARMADA, a multi-robot deployment and adaptation system with human-in-the-loop shared control, featuring an autonomous online failure detection method named FLOAT. Thanks to FLOAT, ARMADA enables paralleled policy rollout and requests human intervention only when necessary, significantly reducing reliance on human supervision. Hence, ARMADA enables efficient acquisition of in-domain data, and leads to more scalable deployment and faster adaptation to new scenarios. We evaluate the performance of ARMADA on four real-world tasks. FLOAT achieves nearly 95% accuracy on average, surpassing prior state-of-the-art failure detection approaches by over 20%. Besides, ARMADA manifests more than 4 $\times$ increase in success rate and greater than 2 $\times$ reduction in human intervention rate over multiple rounds of policy rollout and post-training, compared to previous human-in-the-loop learning methods.

Figures & tables

Fig. 1: Illustration of ARMADA. ARMADA uses the FLOAT failure detector to enable parallel policy rollouts on multiple robots, requesting human intervention only when needed. The resulting online deployment data are then used for policy improvement, forming a scalable deployment-adaptation loop.

Fig. 2: Method overview. FLOAT failure detector conducts real-time OT matching between the policy embeddings of the current rollout and all expert demonstrations, and defines the minimum OT cost as FLOAT index. We thereby calibrate the FLOAT threshold on all successful rollouts. When the FLOAT index of a rollout trajectory exceeds the threshold, we consider it a failure and employ adaptive rewinding based on OT computation, which helps retrace a previous timestep before the scene was disturbed. Our multi-robot system then allocates an idle human operator to the failed robot for intervention.

Fig. 3: Failure detection experiment results. FLOAT achieves nearly 95% accuracy across four tasks, which is an improvement of over 20% compared to state-of-the-art baseline methods. It manifests comparable performance to its variant which further integrates action inconsistency metric, showcasing the effectiveness of FLOAT in detecting various failures.

Fig. 5: Quantitative results of FLOAT and STAC on Fold Towel . Error bars indicate the 5-th and 95-th score quantiles.

Method	TPR (%)	TNR (%)	Accuracy (%)	Weighted Acc. (%)	Step-level TNR (%)
FLOAT	77.96 $\pm$ 7.99	82.46 $\pm$ 11.80	80.21 $\pm$ 1.94	78.20 $\pm$ 2.10	96.80 $\pm$ 3.60
KNN	70.97 $\pm$ 14.10	32.06 $\pm$ 17.10	51.52 $\pm$ 12.30	51.93 $\pm$ 17.80	93.89 $\pm$ 3.00
VIB-ILCtrl	35.69 $\pm$ 5.80	91.51 $\pm$ 1.17	63.60 $\pm$ 3.50	62.76 $\pm$ 9.20	99.18 $\pm$ 0.20

TABLE I: Quantitative comparison with KNN and VIB-ILCtrl. FLOAT achieves the best (Weighted) Accuracy across both tasks, with substantially lower variance.

Fig. 6: Runtime benchmark of FLOAT under varying numbers of expert trajectories and average task lengths.

Panel of Figure 8 — Fig. 7: Qualitative comparison with KNN and VIB-ILCtrl. Top: The robot successfully hangs the mug on the holder. Bottom: The robot fails and drops the mug. FLOAT shows clearer temporal separation between successful and failed executions, leading to more accurate failure detection.

Fig. 8: Success rate over three evaluation rounds. ARMADA exhibits stable progress in success rate, with a more than four-fold increase compared to previous human-in-the-loop learning approach, thanks to our adaptive rewinding mechanism.

Fig. 9: Ablation studies on rewinding mechanism. Comparing training with and without rewinding shows that rewinding substantially improves policy improvement by enabling recovery from hard failure states and yielding more useful intervention data.

Fig. 10: Human intervention rate over three evaluation rounds. ARMADA results in a greater than two-fold reduction in human intervention rate compared to Sirius, showing potential in scalable deployment and adaptation.

Fig. 11: Multi-robot experiments on unseen scenarios. We deploy the pretrained Fold towel policy on Scene A, B, and C (in-domain) for online data collection, and evaluate the post-trained policy on Scene D (out-of-distribution) .

Fig. 12: Success rate on unseen scenarios. ARMADA boosts adaptation to unseen scenarios with paralleled policy deployment on multiple robots.

Fig. 13: Human intervention duration over three deployment rounds. ARMADA scales up collection of human intervention trajectory with more robots in parallel.

Panel of Figure 15 — Fig. 14: Failure analysis of FLOAT. (a) A false-positive sample where FLOAT raises a warning although the mug is eventually hung on the holder successfully. (b) A delayed-detection sample where FLOAT does not identify an early failure promptly and raises a warning after a noticeable delay.

Panel of Figure 16 — Fig. 15: Representative policy failure cases after post-training with ARMADA. Top: Pour marbles into bowl . Bottom: Hang mug on holder . These cases mainly arise from perceptual ambiguity, such as inaccurate estimation of mug rotation or gripper grasp depth, rather than errors that can be fully corrected by DAgger-style behavioral supervision.

Explore similar work

Aug 18, 2026

ORPA: Online Residual Policy Adaptation for Robot Manipulation Control with Human Feedback

Robotic manipulation policies trained via imitation learning, such as Action Chunking with Transformers (ACT), can achieve strong performance under ideal conditions but often remain sensitive to small execution errors and distribution shifts. Correcting these failures typically requires dataset aggregation and full-policy retraining, which is computationally expensive and unsuitable for real-time deployment. In this work, we propose Online Residual Policy Adaptation (ORPA), a framework that enables immediate, feedback-driven correction of robot actions without modifying the underlying policy parameters. ORPA augments a pretrained control policy with a lightweight, feedback-conditioned module that predicts residual adjustments directly in joint space, allowing the system to adapt its behavior at runtime. We evaluate ORPA on a set of precision-sensitive manipulation tasks using the ALOHA platform, demonstrating improvements in success rate and recovery from small perturbations compared to baseline control policies and rule-based inverse kinematics corrections.

Muhammad A. Muttaqien, Tomohiro Motoda, Ryo Hanai +1

Embodied AI Research Team · National Institute of AIST · Tokyo, Japan

Robotic Control Robot Policy Learning Robot Policy Adaptation Robotic Manipulation Residual Policy Learning

Ask PDF

Apr 15, 2026

Failure Identification in Imitation Learning Via Statistical and Semantic Filtering

Imitation learning (IL) policies in robotics deliver strong performance in controlled settings but remain brittle in real-world deployments: rare events such as hardware faults, defective parts, unexpected human actions, or any state that lies outside the training distribution can lead to failed executions. Vision-based Anomaly Detection (AD) methods emerged as an appropriate solution to detect these anomalous failure states but do not distinguish failures from benign deviations. We introduce FIDeL (Failure Identification in Demonstration Learning), a policy-independent failure detection module. Leveraging recent AD methods, FIDeL builds a compact representation of nominal demonstrations and aligns incoming observations via optimal transport matching to produce anomaly scores and heatmaps. Spatio-temporal thresholds are derived with an extension of conformal prediction, and a Vision-Language Model (VLM) performs semantic filtering to discriminate benign anomalies from genuine failures. We also introduce BotFails, a multimodal dataset of real-world tasks for failure detection in robotics. FIDeL consistently outperforms state-of-the-art baselines, yielding +5.30% percent AUROC in anomaly detection and +17.38% percent failure-detection accuracy on BotFails compared to existing methods.

Quentin Rolland, Fabrice Mayran de Chamisso, Jean-Baptiste Mouret

Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France · Inria, CNRS, Université de Lorraine, LORIA, F-54000 Nancy, France · Bleu Robotics, Paris, France

Robot Failure Detection Agent Failure Analysis Anomaly Detection Robot Imitation Learning

Ask PDF

Jul 1, 2026

FAR: Failure-Aware Retry for Test-Time Recovery and Continual Policy Improvement

Robot policies inevitably encounter failures when deployed in real environments. Naive retries often repeat the same mistakes, while many existing recovery methods rely on human intervention. In this paper, we propose Failure-Aware Retry (FAR), a framework that enables robots to learn from previous failures at test time, adapt their behavior accordingly, and eventually complete the task autonomously. FAR combines Failure-Contrastive Preference Adaptation, which constructs preference learning data from failures to steer the policy away from previously unsuccessful behaviors, with lightweight action perturbations during retries to encourage local exploration. We further incorporate successful recovery trajectories into a training loop for continual policy improvement. Experiments in both simulation and real-world manipulation tasks show that FAR substantially improves success rates and robustness, with average gains of 17.6% over the standard diffusion policy in simulation and 11.7% in the real world. In addition, FAR improves data efficiency under both reset and timestep budgets during continual policy improvement by exploiting informative failure cases. Videos and code are available at https://hoar012.github.io/FAR-Project.

Haoran Hao, Shahram Najam Syed, Jeffrey Ichnowski +1

Carnegie Mellon University

Robot Policy Learning Robot Policy Adaptation Robot Skill Learning Robot Failure Recovery Robotic Manipulation Continual RL+1

Ask PDF