When a Robot Kicks a Child

A humanoid robot kicked a child during a public martial arts demonstration in June 2026, sparking debate over liability for AI-powered robot harms. The incident highlights governance gaps as humanoid robot installations are projected to surpass 100,000 units by 2027, raising urgent questions about safety, perception, and control in dynamic human environments.

Global installations of humanoid robots reached approximately 16,000 units in 2025, driven largely by China and early deployments in logistics, manufacturing, and automotive sectors. That figure is set to multiply dramatically: cumulative installations are expected to surpass 100,000 units by 2027. In early June 2026, a video https://www.youtube.com/shorts/BojeUP0 m w from a public robotics demonstration went viral. A humanoid robot, mid-martial arts performance, kicked a young child standing nearby. The child wasn’t seriously hurt, but the moment instantly revived a question that is becoming harder to ignore: Who is liable when an AI-powered robot causes harm? The answer may seem straightforward. One might point to the operator supervising the demonstration, the manufacturer that designed the robot, or the developers responsible for its software https://link.springer.com/article/10.1007/s12027-020-00648-0 . In practice, however, incidents involving AI-powered robots reveal a deeper challenge https://link.springer.com/article/10.1007/s00146-026-02869-2 . As AI increasingly moves from digital environments into physical spaces shared with humans, traditional approaches to liability become harder to apply https://www.europarl.europa.eu/meetdocs/2014 2019/plmrep/COMMITTEES/JURI/DV/2020/01-09/AI-report EN.pdf . The public debate https://www.msn.com/en-in/news/insight/humanoid-robot-s-kick-to-child-in-china-sparks-safety-debate/gm-GM21BCC64F?gemSnapshotKey=GM21BCC64F-snapshot-2&uxmode=ruby surrounding the incident focused primarily on liability after the fact. But that may be the wrong starting point. The more pressing question is why the robot was even capable of hurting a child. Did it have adequate perception? Were its safety limits properly set? Was anyone meaningfully in control? Before we argue about who should pay, we need to ask whether the governance frameworks meant to prevent this kind of thing are working. Before the Demo: Governance Challenges and Risks Beyond Traditional Liability Frameworks Modern humanoid robots operate through interconnected pipelines of sensing, perception, planning, and actuation, https://arxiv.org/pdf/2310.08565v3 processing large volumes of multimodal data in real time. As cyber-physical systems, they rely on interactions among AI models, communication networks, control systems, and hardware components, creating multiple potential attack surfaces. Vulnerabilities in perception, planning, or communication modules can propagate across the robotic stack and ultimately affect physical behavior. Moreover, increasing modularization and interface standardization, exemplified by initiatives such as the Hardware Robot Information Model HRIM , https://arxiv.org/pdf/2310.08565v3 are improving the interoperability and scalability of robotic systems. However, they also complicate accountability and cybersecurity, as failures in one component may cascade across interconnected modules developed by different actors. Future governance frameworks may therefore need to focus on the safety, traceability, and trustworthiness of entire embodied AI systems rather than individual AI models alone. If a public robotics demonstration reveals the risks associated with deploying AI systems in dynamic human environments, those risks become even more significant in healthcare. Hospitals and clinical settings provide a particularly useful lens through which to examine these challenges. AI-powered robots are increasingly used in surgery https://link.springer.com/article/10.1007/s11701-024-01867-0 , diagnostics https://onlinelibrary.wiley.com/doi/full/10.1111/cas.14377 , rehabilitation https://journals.sagepub.com/doi/full/10.1155/2014/563062 , and clinical decision support https://www.acpjournals.org/doi/full/10.7326/0003-4819-157-1-201207030-00450 , often operating in contexts where human vulnerability is heightened and the margin for error is exceptionally small. Unlike traditional medical devices that follow predefined instructions, AI-enabled systems can adapt to changing circumstances, learn from data, and respond to situations that were not explicitly anticipated by their developers. As a result https://pmc.ncbi.nlm.nih.gov/articles/PMC10879008/ , questions of safety, governance, and liability become considerably more complex than under traditional models of medical technology. A recent study https://link.springer.com/article/10.1007/s00146-026-02869-2 examining the regulatory and liability implications of AI-powered medical robots in the European Union revealed significant uncertainty regarding the legal frameworks governing these technologies. Drawing on surveys of 50 medical professionals across 20 countries and 19 specialties, expert interviews, and normative legal analysis, the study found that more than half of respondents were unaware of the legal frameworks applicable to AI and robotics in healthcare, while only 1.7% considered themselves well informed. The findings also revealed considerable uncertainty regarding liability. Nearly 60% of respondents questioned whether existing EU liability rules adequately address harms caused by AI-powered robots, while almost 80% called for clearer regulatory guidance before such systems are more widely deployed. Taken together, these findings suggest that the challenge is not merely technological. As AI systems become increasingly autonomous and embedded in environments where human wellbeing is at stake, legal clarity, effective governance, and meaningful accountability may prove just as important as technical performance. Why Traditional Liability Models Fall Short Traditional liability frameworks file:///C:/Users/u0149483/Downloads/1999IntlFinLRev25.pdf were designed for a world in which harm can be traced to a clearly identifiable actor. A physician makes a mistake. A manufacturer produces a defective product. Responsibility follows a relatively linear path. AI-powered robots break this logic https://pspac.info/index.php/dlbh/article/view/247/197 . Legal scholars have long anticipated this moment. Writing in 2016 https://watermark02.silverchair.com/nclr 2016 19 3 412.pdf?token=AQECAHi208BE49Ooan9kkhW Ercy7Dm3ZL 9Cf3qfKAc485ysgAAA0gwggNEBgkqhkiG9w0BBwagggM1MIIDMQIBADCCAyoGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMArAlS11P8LKmi2O5AgEQgIIC-x4Hc0ykxALMeK3RxW-DCxwZqKPH95Gc 5MCEPmBGBzCzIcmRuCIIIF38cA1SL0 3GYaa7Av3V7cTiC 8O-V5utYIvTmkbltWebdaaGeAIVOAtkK1wlVtRvKkTGYWJ9AFF6HudJ5fsmDzpAz0X-vTdrc6 C57iMXuHmT27KfXmLxUQkS5arckeko Q7G2xLU6xYzD7wzX13oaRwG3v4BuVeGSlI axiJSykDTV7kGg6EG4t5rNI 2oq9iHK75CvfqOmsWp8Ddi4usAgA4FCOxjR53T1x0JfExcBLUIaoUIxYcxV3SJNOf KnY8g0HynQSD2wQXdB6ye6xyGHbVmKg-gxFQNlk1pYi2dbXGYKAhpAP7g-RfTTRT6SFTJE2DXJ9v QvfKfcCBz-xSKlEZdHZqCVKv-eFoFDR5JUrxRRy2zbq9pnKh8IBlPSQPCy1qD-eIs4oZlLCDUORqQyK-LE9mPv xZE2m-QKgTHewTxxSHIyTp07oMZKWA7fZJTHaw0tbFzmQVN7CgTDyBKAAGoBRhSOTXpESnx7AwHlh8Ya4Ts5MiXFxDdHJpzHHDWKEAKTQ8TA5-YTHxvtBlsq XFTrNltuZ2Bto03xlyzXhxJP1-e-JtwlCHelA7aRmwLh9UIIG1Kz7uW5se2V1jHkIN3ZW5AN4AXnODYb3w2s--7jwGZdTWKqXmTERvNZGpH7CXFu0T0NaLwb1HSU2KHoQ20WlXpgvNLeC95vfRJHWtRoj6N-FFkVRcoZHN1vT-Yp02yq2LbYINYsObolq0nQAj72muE8EQZWlI2naZamZluJdQzOMa-oeb33ilnD0kDNO0i79kB8EerX6muVGvnLHqYi2Jeosx3V2VqQABXUmkHM89r5BG--nfpBjxQcvyQpnmnxKVEZS S8Havgst78I1pCEe9YQdawprPWeizw1uu-blF4xeeM qkfAf1sgAy CJzE7m3M 2e zMlV8gyWgeRWHtOZN-FDoxK6ltpqc4qX-VCqHfRYli30giTk , Gless, Silverman, and Weigend argued that as long as robots lack moral self-awareness, liability must fall on the humans who design, program, and deploy them, and that the unpredictability of an AI system is not a defense, but precisely the source of the duty of care. Therefore, liability is rarely concentrated in a single actor https://link.springer.com/article/10.1007/s00146-026-02869-2 . Instead, it emerges from a chain of interconnected decisions made across the lifecycle of the AI-powered robot. The study https://link.springer.com/article/10.1007/s00146-026-02869-2 identifies eight key factors that influence liability in this context, including the explainability of AI decisions, the quality and bias of training data, cybersecurity vulnerabilities, language barriers, the degree of automation versus autonomy, and regulatory variability. Each of these complicates the attribution of fault in ways that existing doctrine was not designed to handle. A Link-in-the-Chain Liability Model To address this complexity, a link-in-the-chain liability model has been proposed in literature https://link.springer.com/article/10.1007/s00146-026-02869-2 . Rather than assigning responsibility exclusively to a single actor, the model recognises that AI systems are the product of multiple interconnected decisions made throughout their lifecycle. Designers, developers, manufacturers, deployers, operators, and organisations integrating AI into real-world environments all contribute to how these systems function. Therefore, liability should be assessed in light of each actor’s degree of control, their ability to foresee risks, and whether they fulfilled their respective obligations. The model integrates two dimensions https://link.springer.com/article/10.1007/s00146-026-02869-2 . The ex ante dimension focuses on preventing harm through regulatory compliance: ensuring that safety is designed in from the outset, not retrofitted after an incident. The ex post dimension addresses compensation and accountability when harm nonetheless occurs. Neither dimension alone is sufficient. Together, they offer a more coherent framework for governing AI-powered robots than either medical malpractice or product liability regimes can provide on their own. From this perspective, the AI Act https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng can be understood as an attempt to move beyond purely compensatory models of liability. Rather than asking who should be responsible after harm occurs, the AI Act seeks to embed responsibility throughout the lifecycle of AI systems by imposing obligations on providers, deployers, and other actors involved in their development and use. The AI Act, the Digital Omnibus, and What Changes — and What Doesn’t The EU AI Act reflects a broader shift from ex post liability towards ex ante governance. For high-risk AI systems, it imposes obligations relating to risk management, data governance, technical documentation, record-keeping, transparency, robustness, and human oversight. While the Act does not prescribe specific technical solutions, incidents involving humanoid robots raise important questions about whether existing safeguards are sufficient when AI systems operate in close proximity to humans. https://www.technologyreview.com/2025/06/11/1118519/humanoids-safety-rules/ However, the EU regulatory picture has recently become more complex. In November 2025 https://digital-strategy.ec.europa.eu/en/library/digital-omnibus-ai-regulation-proposal , the European Commission proposed the Digital Omnibus on AI as part of its broader simplification agenda https://commission.europa.eu/law/law-making-process/better-regulation/simplification-implementation-and-enforcement/simplification en , with a provisional political agreement reached in May 2026. While the Omnibus does not alter the substantive obligations of the AI Act, https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:52025PC0837 it adjusts implementation timelines and simplifies certain compliance pathways, particularly for smaller providers. The rationale is largely practical. Harmonised technical standards remain under development https://digital-strategy.ec.europa.eu/en/policies/ai-act-standardisation , and several Member States are still in the process of establishing and operationalising their national supervisory structures. However, the postponement of certain compliance obligations also raises important governance questions https://www.moorelaw.be/en/news/ai-act-update-postponement-of-highrisk-obligations-and-clarification-of-transparency . As increasingly capable AI systems continue to enter real-world environments, regulatory delays may create a gap between technological deployment and effective oversight. It may simplify https://www.delorscentre.eu/en/publications/detail/publication/the-eus-digital-and-ai-omnibus certain cybersecurity and data-related rules, but the AI Act’s core requirements, risk management, transparency, traceability, robustness, and human oversight, remain intact. Therefore, the broader challenge lies not in the substance of the rules, but in ensuring that they can be implemented consistently and effectively across different sectors and use cases. Nor does the Omnibus resolve the wider problem of regulatory fragmentation. AI systems often operate at the intersection of multiple legal regimes, including data protection, product safety, product liability, cybersecurity, and sector-specific legislation. As a result, questions of liability frequently extend beyond https://link.springer.com/article/10.1007/s00146-026-02869-2 the AI Act itself, requiring actors to navigate a complex and evolving regulatory landscape. Regulatory frameworks could impose requirements that mandate the retention of operational data, AI model versions, system updates, and decision logs throughout a robot’s lifecycle. In addition, continuous recording of robot activities could provide an auditable trail for incident investigation and regulatory oversight. As humanoid robots become more autonomous, questions of what data should be retained, where it should be stored, and for how long may become central governance issues. Safety-by-Design, Not Liability-by-Default The video illustrates a fundamental asymmetry in discussions about AI governance https://thefuturesociety.org/wp-content/uploads/2026/05/The Case for Cross-Border AI Incident Infrastructure.pdf . Public attention tends to focus on liability only after harm has occurred. Once a child has been injured, questions immediately arise about who should be held responsible. However, effective governance begins much earlier than the moment of deployment. As Ribeiro et al. argue https://arxiv.org/abs/2505.23417 , governance is not limited to legal compliance, it encompasses the design of institutional arrangements that operationalize ethical principles throughout the entire AI lifecycle, from development through to real-world use. This reflects a broader shift from liability-by-default towards safety-by-design. Degeorges and Sziebig https://ieeexplore.ieee.org/abstract/document/9576407?casa token=XSJlNxI2uMsAAAAA:x5s9pL7fu0wjp8Fq1EYXDtpO0x7cQp9P2EF5IX1WHIboXlITjBdO-50L6xgQGR020UfboEze make this concrete in the context of human-robot collaboration: reactive safety mechanisms that respond only to proximity are not enough . What is needed is a proactive architecture in which the system continuously models its environment and anticipates potential harm before it occurs. The question, then, is not only who is responsible after something goes wrong, but whether the system was ever designed to prevent it in the first place. This does not render liability irrelevant. When harm occurs, mechanisms for accountability and compensation remain essential. But video-shot incidents expose the limits of a purely reactive approach, and they expose something more specific: the particular risks that arise when AI is not just software running in the background, but a physical system acting in the world . Embodied AI systems move, exert force, and share space with people. When they fail, the consequences are immediate and corporeal. Liability determines who bears responsibility after harm has materialized, and safety-by-design seeks to reduce the likelihood of that harm occurring at all. From this perspective, liability is not a question that arises only after an accident. It extends across the entire system lifecycle. The actors involved in designing, developing, deploying, and operating embodied AI systems each shape the risks those systems may create risks that, unlike a biased algorithm or flawed recommendation, can land with physical force on a child standing nearby. Effective governance therefore requires both ex ante safeguards aimed at preventing harm and ex post mechanisms capable of allocating responsibility when prevention fails. The challenge for policymakers is therefore not simply to determine who should be liable when robots cause harm, but to ensure that the conditions that make such harm possible are identified and addressed before deployment. Acknowledgements This blog post draws on findings disseminated through the HEREDITARY project https://www.law.kuleuven.be/citip/en/research/projects/ongoing/hereditary , including research published in AI & Society . It was developed in the context of the IEEE CIS Task Force on AI Governance, Regulation and Compliance. https://cis.taskforce.ieee.org/airc/