On 5 January 2026 Nvidia announced new open models, frameworks and AI infrastructure for physical AI at CES 2026 in Las Vegas, and showcased robots built by global partners. According to the company, the technology is intended to accelerate the entire robotics development lifecycle, including building “expert generalist” robots that can quickly learn multiple tasks.
What physical AI means for industry
Physical AI refers to model systems that let machines understand and interact with the physical world; it is the technical foundation of embodied intelligence. Unlike pure-software generative AI, physical AI must close the perception-decision-execution loop and is ultimately delivered as a hardware product.
This shift suggests that the focus of AI governance may extend from content compliance to product compliance. Once a model is embedded in a robot, an autonomous vehicle or industrial equipment, what is involved is not only synthetic-content labelling but also product safety, allocation of liability and export controls.
Three compliance issues
First, the boundaries of training-data collection. Physical AI depends on real-world visual, motion and sensor data. Such data often includes identifiable images of individuals and industrial scene information that may constitute important data. Whether the authorisation chain at collection is complete will determine whether the model can be used lawfully. See our data compliance service.
Second, downstream constraints of open-source licences. Open-source models lower the R&D barrier, but licence conditions for commercial use, attribution obligations and derivative-work rules differ. See our notes on the AI open-source ecosystem consensus and Kimi K2.6’s compliance arithmetic.
Third, allocation of product liability. Where a robot causes personal injury or property damage, how liability is divided among the model provider, the integrator and the user is not yet settled. The EU AI Act has set out market-placement duties for high-risk systems; see our note on the US AI model review executive order.
Three points for Chinese companies
First, if importing an overseas physical-AI foundation model, assess whether it is affected by export controls or investment security review — see our note on the NDRC’s block of the Manus acquisition.
Second, if developing embodied-AI products, build a compliance file for data collection early in R&D rather than remediating at the production stage.
Third, if products target both domestic and overseas markets, assess them separately against each jurisdiction’s high-risk classification.
The industrial window for physical AI is opening, but the rules to match it still lag. For companies, first-mover advantage lies not only in technology but in whether compliance design is built into the product architecture from the outset.
This article is based on Nvidia’s CES 2026 announcements and public reporting. It is for general reference only and does not constitute legal advice.