{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,2]],"date-time":"2026-09-02T09:22:53Z","timestamp":1788340973479,"version":"build-2803163510"},"reference-count":51,"publisher":"MDPI AG","issue":"9","license":[{"start":{"date-parts":[[2026,9,2]],"date-time":"2026-09-02T00:00:00Z","timestamp":1788307200000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"name":"Smart Tower Co., Ltd."},{"name":"Research on General Tower Intelligent Agent"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["Entropy"],"abstract":"<jats:p>Vision\u2013Language\u2013Action (VLA) policies remain brittle under modest distribution shift. On LIBERO-Plus, contemporary models that solve clean tasks at high rates can fall below 30% success when the camera\u2019s viewpoint or the robot\u2019s initial pose is perturbed. Most training-free test-time remedies address this problem through the image stream, for example, by augmenting, purifying, or selecting visual observations. In our controlled evaluation, this family of methods improves mean success by only about three points and leaves the robot-initial-state failure largely unresolved. This paper studies the failure at the level of input streams. A VLA receives visual tokens, a proprioceptive state token, and language tokens; different perturbations can move different streams away from their training manifold. In particular, the robot-initial-state perturbation directly shifts the proprioceptive token; therefore, image-space interventions have limited leverage. We introduce Gated Per-Stream Manifold Restoration (G-PSMR), a training-free wrapper for a frozen policy. For each stream, a lightweight gate detects off-manifold inputs and applies a stream-specific restoration before the policy forward pass. We instantiate the framework with entropy-gated visual consensus and gated relative-orientation debiasing, which preserves the within-episode orientation trajectory. In the original 280-episode paired evaluation, the joint method improves total success by +5.3 points compared with a +3.2 image-only gain and raises the most fragile factor from 20% to 38%. On 1120 previously unevaluated, manifest-disjoint task instances, the state restoration improves robot-initial-state success from 23.8% to 28.7%; a separate prospectively specified confirmation on 600 new gate-active instances yields 26.3%\u219232.2% (+5.8 points; 95% CI [+3.2,+8.5]; p&lt;0.001). Together, the original cross-stream results and two independent state-stream evaluations support the central principle of matching the restoration to the input stream carrying the shift.<\/jats:p>","DOI":"10.3390\/e28090975","type":"journal-article","created":{"date-parts":[[2026,9,2]],"date-time":"2026-09-02T08:56:01Z","timestamp":1788339361000},"page":"975","update-policy":"https:\/\/doi.org\/10.3390\/mdpi_crossmark_policy","source":"Crossref","is-referenced-by-count":0,"title":["Restoring the Right Stream: Training-Free OOD Robustness for Vision\u2013Language\u2013Action Policies"],"prefix":"10.3390","volume":"28","author":[{"given":"Zhiyin","family":"Yan","sequence":"first","affiliation":[{"name":"Smart Tower Co., Ltd., Beijing 100195, China"}],"role":[{"vocabulary":"crossref","role":"author"},{"role":"conceptualization","vocabulary":"credit"},{"role":"methodology","vocabulary":"credit"},{"role":"software","vocabulary":"credit"},{"role":"validation","vocabulary":"credit"},{"role":"formal-analysis","vocabulary":"credit"},{"role":"writing-review-editing","vocabulary":"credit"},{"role":"writing-original-draft","vocabulary":"credit"},{"role":"project-administration","vocabulary":"credit"},{"role":"data-curation","vocabulary":"credit"},{"role":"resources","vocabulary":"credit"},{"role":"supervision","vocabulary":"credit"},{"role":"visualization","vocabulary":"credit"},{"role":"investigation","vocabulary":"credit"}]},{"given":"Gang","family":"Shen","sequence":"additional","affiliation":[{"name":"Smart Tower Co., Ltd., Beijing 100195, China"}],"role":[{"vocabulary":"crossref","role":"author"},{"role":"visualization","vocabulary":"credit"},{"role":"conceptualization","vocabulary":"credit"},{"role":"methodology","vocabulary":"credit"},{"role":"software","vocabulary":"credit"},{"role":"validation","vocabulary":"credit"},{"role":"funding-acquisition","vocabulary":"credit"}]}],"member":"1968","published-online":{"date-parts":[[2026,9,2]]},"reference":[{"key":"ref_1","first-page":"2679","article-title":"OpenVLA: An Open-Source Vision-Language-Action Model","volume":"Volume 270","author":"Kim","year":"2025","journal-title":"Proceedings of the Conference on Robot Learning (CoRL), Munich, Germany, 6\u20139 November 2024"},{"key":"ref_2","doi-asserted-by":"crossref","unstructured":"Kim, M.J., Finn, C., and Liang, P. (2025, January 21\u201325). Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. Proceedings of the Robotics: Science and Systems (RSS), Los Angeles, CA, USA.","DOI":"10.15607\/RSS.2025.XXI.017"},{"key":"ref_3","doi-asserted-by":"crossref","unstructured":"Fei, S., Wang, S., Shi, J., Dai, Z., Cai, J., Qian, P., Ji, L., He, X., Zhang, S., and Fei, Z. (2025). LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models. arXiv.","DOI":"10.15607\/RSS.2026.XXII.097"},{"key":"ref_4","unstructured":"Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., and Wahid, A. (2023, January 6\u20139). RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. Proceedings of the Conference on Robot Learning (CoRL), Atlanta, GA, USA."},{"key":"ref_5","unstructured":"Octo Model Team, Ghosh, D., Walke, H., Pertsch, K., Black, K., Mees, O., Dasari, S., Hejna, J., Kreiman, T., and Xu, C. (2024, January 15\u201319). Octo: An Open-Source Generalist Robot Policy. Proceedings of the Robotics: Science and Systems (RSS), Delft, The Netherlands."},{"key":"ref_6","doi-asserted-by":"crossref","unstructured":"Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., and Ichter, B. (2025, January 21\u201325). \u03c00: A Vision-Language-Action Flow Model for General Robot Control. Proceedings of the Robotics: Science and Systems (RSS), Los Angeles, CA, USA.","DOI":"10.15607\/RSS.2025.XXI.010"},{"key":"ref_7","doi-asserted-by":"crossref","unstructured":"Pertsch, K., Stachowicz, K., Ichter, B., Driess, D., Nair, S., Vuong, Q., Mees, O., Finn, C., and Levine, S. (2025, January 21\u201325). FAST: Efficient Action Tokenization for Vision-Language-Action Models. Proceedings of the Robotics: Science and Systems (RSS), Los Angeles, CA, USA.","DOI":"10.15607\/RSS.2025.XXI.012"},{"key":"ref_8","doi-asserted-by":"crossref","unstructured":"Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., and Stone, P. (2023, January 10\u201316). LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. Proceedings of the NeurIPS Datasets and Benchmarks Track, New Orleans, LA, USA.","DOI":"10.52202\/075280-1939"},{"key":"ref_9","doi-asserted-by":"crossref","first-page":"FSE073","DOI":"10.1145\/3729343","article-title":"VLATest: Testing and Evaluating Vision-Language-Action Models for Robotic Manipulation","volume":"2","author":"Wang","year":"2025","journal-title":"Proc. ACM Softw. Eng."},{"key":"ref_10","unstructured":"Guo, J., Wu, Z., Tu, C., Ma, Y., Kong, X., Liu, Z., Ji, J., Zhang, S., Chen, Y., and Chen, K. (2026, January 23\u201327). On Robustness of Vision-Language-Action Model against Multi-Modal Perturbations. Proceedings of the International Conference on Learning Representations (ICLR), Rio de Janeiro, Brazil."},{"key":"ref_11","unstructured":"Atreya, P., Pertsch, K., Lee, T., Kim, M.J., Jain, A., Kuramshin, A., Neary, C., Hu, E.S., Arora, K., and Ellis, K. (2025, January 27\u201330). RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies. Proceedings of the Conference on Robot Learning (CoRL), Seoul, Republic of Korea."},{"key":"ref_12","unstructured":"de Haan, P., Jayaraman, D., and Levine, S. (2019, January 8\u201314). Causal Confusion in Imitation Learning. Proceedings of the NeurIPS, Vancouver, BC, Canada."},{"key":"ref_13","unstructured":"Zhao, J., Lu, W., Zhang, D., Liu, Y., Liang, Y., Zhang, T., Cao, Y., Xie, J., Hu, Y., and Wang, S. (2025). Do You Need Proprioceptive States in Visuomotor Policies?. arXiv."},{"key":"ref_14","unstructured":"Li, Z., Li, Y., Jiang, J.-J., Zhou, L., Zhang, T., Yin, J., Lin, M., Wei, Y.-L., and Zheng, W.-S. (2026). ReViP: Mitigating False Completion in Vision-Language-Action Models with Vision-Proprioception Rebalance. arXiv."},{"key":"ref_15","doi-asserted-by":"crossref","unstructured":"Huang, Y., Peng, P., Zhao, Y., Chen, G., and Tian, Y. (2024, January 10\u201315). Seek Commonality but Preserve Differences: Dissected Dynamics Modeling for Multi-modal Visual RL. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), Vancouver, BC, Canada.","DOI":"10.52202\/079017-0053"},{"key":"ref_16","doi-asserted-by":"crossref","unstructured":"Zhang, L., Chen, G., Shi, Y., Zhang, Q., Tan, M., Wang, Y., Tian, Y., and Huang, T. (2020, January 6\u20138). Anonymous Model Pruning for Compressing Deep Neural Networks. Proceedings of the 2020 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR), Shenzhen, China.","DOI":"10.1109\/MIPR49039.2020.00040"},{"key":"ref_17","doi-asserted-by":"crossref","unstructured":"Luo, L., Chen, G., Zheng, X., Dai, Y., Zou, Y., and Tian, Y. (2026, January 20\u201327). Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning. Proceedings of the AAAI Conference on Artificial Intelligence, Singapore.","DOI":"10.1609\/aaai.v40i9.37709"},{"key":"ref_18","doi-asserted-by":"crossref","unstructured":"Zeng, Y., Wu, H., Nie, W., Chen, G., Zheng, X., Shen, Y., Peng, J., Tian, Y., and Ji, R. (2025, January 19\u201325). From Objects to Events: Unlocking Complex Visual Understanding in Object Detectors via LLM-guided Symbolic Reasoning. Proceedings of the IEEE\/CVF International Conference on Computer Vision (ICCV), Honolulu, HI, USA.","DOI":"10.1109\/ICCV51701.2025.02260"},{"key":"ref_19","doi-asserted-by":"crossref","unstructured":"Song, B., Peng, J., Zhang, Y., Chen, G., Yang, F., and Guo, J. (2026, January 20\u201327). KTV: Keyframes and Key Tokens Selection for Efficient Training-Free Video LLMs. Proceedings of the AAAI Conference on Artificial Intelligence, Singapore.","DOI":"10.1609\/aaai.v40i11.37862"},{"key":"ref_20","unstructured":"Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A.A., and Hardt, M. (2020, January 13\u201318). Test-Time Training with Self-Supervision for Generalization under Distribution Shifts. Proceedings of the ICML, Virtual."},{"key":"ref_21","unstructured":"Wang, D., Shelhamer, E., Liu, S., Olshausen, B., and Darrell, T. (2021, January 3\u20137). Tent: Fully Test-time Adaptation by Entropy Minimization. Proceedings of the ICLR, Virtual."},{"key":"ref_22","unstructured":"Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., and Tan, M. (2022, January 17\u201323). Efficient Test-Time Model Adaptation without Forgetting. Proceedings of the ICML, Baltimore, MD, USA."},{"key":"ref_23","unstructured":"Zhang, M., Levine, S., and Finn, C. (December, January 28). MEMO: Test Time Robustness via Adaptation and Augmentation. Proceedings of the NeurIPS, New Orleans, LA, USA."},{"key":"ref_24","doi-asserted-by":"crossref","unstructured":"Shanmugam, D., Blalock, D., Balakrishnan, G., and Guttag, J. (2021, January 11\u201317). Better Aggregation in Test-Time Augmentation. Proceedings of the ICCV, Virtual.","DOI":"10.1109\/ICCV48922.2021.00125"},{"key":"ref_25","unstructured":"Kwok, J., Agia, C., Sinha, R., Foutter, M., Li, S., Stoica, I., Mirhoseini, A., and Pavone, M. (2025, January 27\u201330). RoboMonkey: Scaling Test-Time Sampling and Verification for Vision-Language-Action Models. Proceedings of the Conference on Robot Learning (CoRL), Seoul, Republic of Korea."},{"key":"ref_26","unstructured":"Jang, S., Kim, D., Kim, C., Kim, Y., and Shin, J. (2026, January 23\u201327). Verifier-free Test-Time Sampling for Vision-Language-Action Models. Proceedings of the International Conference on Learning Representations (ICLR), Rio de Janeiro, Brazil."},{"key":"ref_27","unstructured":"Lin, J., Taherin, A., Akbari, A., Akbari, A., Lu, L., Chen, G., Padir, T., Yang, X., Chen, W., and Li, Y. (2025). VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting. arXiv."},{"key":"ref_28","first-page":"689","article-title":"Unpacking Failure Modes of Generative Policies: Runtime Monitoring of Consistency and Progress","volume":"Volume 270","author":"Agia","year":"2025","journal-title":"Proceedings of the Conference on Robot Learning (CoRL), Munich, Germany, 6\u20139 November 2024"},{"key":"ref_29","doi-asserted-by":"crossref","unstructured":"Xu, C., Nguyen, T.K., Dixon, E., Rodriguez, C., Miller, P., Lee, R., Shah, P., Ambrus, R., Nishimura, H., and Itkina, M. (2025, January 21\u201325). Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies. Proceedings of the Robotics: Science and Systems (RSS), Los Angeles, CA, USA.","DOI":"10.15607\/RSS.2025.XXI.073"},{"key":"ref_30","unstructured":"Gu, Q., Ju, Y., Sun, S., Gilitschenski, I., Nishimura, H., Itkina, M., and Shkurti, F. (2025, January 2\u20137). SAFE: Multitask Failure Detection for Vision-Language-Action Models. Proceedings of the NeurIPS, San Diego, CA, USA. Mexico City, Mexico, 30 November\u20135 December 2025."},{"key":"ref_31","unstructured":"Ganai, M., Sinha, R., Agia, C., Morton, D., Di Lillo, L., and Pavone, M. (2025, January 27\u201330). Real-Time Out-of-Distribution Failure Prevention via Multi-Modal Reasoning. Proceedings of the Conference on Robot Learning (CoRL), Seoul, Republic of Korea."},{"key":"ref_32","unstructured":"Turner, A.M., Thiergart, L., Leech, G., Udell, D., Vazquez, J.J., Mini, U., and MacDiarmid, M. (2023). Steering Language Models with Activation Engineering. arXiv."},{"key":"ref_33","doi-asserted-by":"crossref","unstructured":"Li, K., Patel, O., Vi\u00e9gas, F., Pfister, H., and Wattenberg, M. (2023, January 10\u201316). Inference-Time Intervention: Eliciting Truthful Answers from a Language Model. Proceedings of the NeurIPS, New Orleans, LA, USA.","DOI":"10.52202\/075280-1797"},{"key":"ref_34","doi-asserted-by":"crossref","unstructured":"Rimsky, N., Gabrieli, N., Schulz, J., Tong, M., Hubinger, E., and Turner, A.M. (2024, January 11\u201316). Steering Llama 2 via Contrastive Activation Addition. Proceedings of the ACL, Bangkok, Thailand.","DOI":"10.18653\/v1\/2024.acl-long.828"},{"key":"ref_35","unstructured":"H\u00e4on, B., Stocking, K.C., Chuang, I., and Tomlin, C. (2025, January 27\u201330). Mechanistic Interpretability for Steering Vision-Language-Action Models. Proceedings of the Conference on Robot Learning (CoRL), Seoul, Republic of Korea."},{"key":"ref_36","unstructured":"Ho, J., and Salimans, T. (2022). Classifier-Free Diffusion Guidance. arXiv."},{"key":"ref_37","unstructured":"Cao, J., Huang, Y., Guo, H., Zhang, Q., Zhang, R., Mai, W., Nan, M., Wang, J., Cheng, H., and Sun, J. (2026, January 23\u201327). Compose Your Policies! Improving Diffusion-based or Flow-based Robot Policies via Test-time Distribution-level Composition. Proceedings of the International Conference on Learning Representations (ICLR), Rio de Janeiro, Brazil."},{"key":"ref_38","unstructured":"Nie, W., Guo, B., Huang, Y., Xiao, C., Vahdat, A., and Anandkumar, A. (2022, January 17\u201323). Diffusion Models for Adversarial Purification. Proceedings of the ICML, Baltimore, MD, USA."},{"key":"ref_39","doi-asserted-by":"crossref","unstructured":"Gao, J., Zhang, J., Liu, X., Darrell, T., Shelhamer, E., and Wang, D. (2023, January 17\u201324). Back to the Source: Diffusion-Driven Adaptation to Test-Time Corruption. Proceedings of the CVPR, Vancouver, BC, Canada.","DOI":"10.1109\/CVPR52729.2023.01134"},{"key":"ref_40","doi-asserted-by":"crossref","unstructured":"Hancock, A.J., Ren, A.Z., and Majumdar, A. (2024). Run-Time Observation Interventions Make Vision-Language-Action Models More Visually Robust. arXiv.","DOI":"10.1109\/ICRA55743.2025.11128017"},{"key":"ref_41","unstructured":"Abouzeid, A., Mansour, M., Sun, Q., Sun, Z., and Song, D. (2025). GeoAware-VLA: Implicit Geometry Aware Vision-Language-Action Model. arXiv."},{"key":"ref_42","unstructured":"Heo, H., Woo, S., Kim, S.M., Kim, J., Lee, J., Lee, Y., and Kim, Y.M. (2026). AnyCamVLA: Zero-Shot Camera Adaptation for Viewpoint Robust Vision-Language-Action Models. arXiv."},{"key":"ref_43","unstructured":"Kaba, S.O., Mondal, A.K., Zhang, Y., Bengio, Y., and Ravanbakhsh, S. (2023, January 23\u201329). Equivariance with Learned Canonicalization Functions. Proceedings of the ICML, Honolulu, HI, USA."},{"key":"ref_44","doi-asserted-by":"crossref","first-page":"983","DOI":"10.1090\/jams\/852","article-title":"Testing the Manifold Hypothesis","volume":"29","author":"Fefferman","year":"2016","journal-title":"J. Am. Math. Soc."},{"key":"ref_45","unstructured":"Liu, W., Wang, X., Owens, J., and Li, Y. (2020, January 6\u201312). Energy-based Out-of-distribution Detection. Proceedings of the NeurIPS, Virtual."},{"key":"ref_46","doi-asserted-by":"crossref","unstructured":"Chen, G., Horstmann, K., Wang, Z., and You, F. (2025, January 11\u201315). Automated Essential Concept Discovery for Few-Shot Out-of-Distribution Detection. Proceedings of the IEEE\/CVF Conference on Computer Vision and Pattern Recognition Workshops, Nashville, TN, USA.","DOI":"10.1109\/CVPRW67362.2025.00381"},{"key":"ref_47","unstructured":"Zhang, Z., Chen, G., Zou, Y., Li, Y., and Li, R. (November, January 28). Learning Unknowns from Unknowns: Diversified Negative Prototypes Generator for Few-Shot Open-Set Recognition. Proceedings of the 32nd ACM International Conference on Multimedia, Melbourne, Australia."},{"key":"ref_48","unstructured":"Xiao, D., Chen, G., Peng, P., Huang, Y., Zhao, Y., Dai, Y., and Tian, Y. (2025, January 13\u201319). When Every Millisecond Counts: Real-Time Anomaly Detection via the Multimodal Asynchronous Hybrid Network. Proceedings of the International Conference on Machine Learning (ICML), Vancouver, BC, Canada."},{"key":"ref_49","unstructured":"Samangouei, P., Kabkab, M., and Chellappa, R. (May, January 30). Defense-GAN: Protecting Classifiers Against Adversarial Attacks Using Generative Models. Proceedings of the ICLR, Vancouver, BC, Canada."},{"key":"ref_50","doi-asserted-by":"crossref","unstructured":"Stutz, D., Hein, M., and Schiele, B. (2019, January 16\u201320). Disentangling Adversarial Robustness and Generalization. Proceedings of the CVPR, Long Beach, CA, USA.","DOI":"10.1109\/CVPR.2019.00714"},{"key":"ref_51","doi-asserted-by":"crossref","unstructured":"Reichlin, A., Marchetti, G.L., Yin, H., Ghadirzadeh, A., and Kragic, D. (2022, January 23\u201327). Back to the Manifold: Recovering from Out-of-Distribution States. Proceedings of the IROS, Kyoto, Japan.","DOI":"10.1109\/IROS47612.2022.9981315"}],"container-title":["Entropy"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/www.mdpi.com\/1099-4300\/28\/9\/975\/pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,9,2]],"date-time":"2026-09-02T08:58:04Z","timestamp":1788339484000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.mdpi.com\/1099-4300\/28\/9\/975"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,9,2]]},"references-count":51,"journal-issue":{"issue":"9","published-online":{"date-parts":[[2026,9]]}},"alternative-id":["e28090975"],"URL":"https:\/\/doi.org\/10.3390\/e28090975","relation":{},"ISSN":["1099-4300"],"issn-type":[{"value":"1099-4300","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,9,2]]}}}