Embodied AI Safety: SAE Panel Says Execution Beats Model Power

A panel that put model hype second

The robotics industry's 2026 has been loud about scale: fleet counts, factory capacity and new versions of foundation models. A white paper drawn from an SAE World Congress 2026 panel called Embodied AI in Action takes a quieter position. It argues that the companies that win with physical AI will be the ones that can field it safely, not the ones that build the most capable model.

The paper was posted to arXiv's robotics category on May 11, 2026. It summarizes a session with experts from automotive engineering, robotics, AI and safety engineering, and it says the panel broadly agreed that long-term success depends as much on safe and trustworthy deployment as on gains in AI capability1. The same document is also on ResearchGate, where it is framed as practical guidance for executives, policymakers and technical leaders2.

That message lands while humanoid robot makers are making very different claims about how far they have gotten. Comparing the panel's framework with this year's deployment reporting shows why the panel's caution matters.

What the panel actually said

The paper's main point is that embodied AI is a systems problem, not just a software problem. The panel said safe deployment depends on sensors, hardware robustness, operating limits, human interaction, communications, cybersecurity and organizational processes, along with the algorithms1. It defined embodied AI simply as AI that can do something physical, such as driving, lifting, navigating or manipulating objects. Once a machine does those things, safety and reliability become central concerns1.

Several other themes recur in the paper:

• Benchmarks are not readiness. Panelists with machine-learning backgrounds warned that a model can test well and still fail on rare scenarios, degraded sensors or unexpected human behavior. They said complete systems should be evaluated, not isolated algorithms1.
• Keep the old safety tools. Safety engineers argued that hazard analysis, requirements traceability, failure-mode assessment and layered mitigations still apply. They need to be extended to cover data dependencies, probabilistic behavior and performance drift over time1.
• Rare events matter more than dataset size. The panel called for data strategies built around operational relevance rather than sheer volume. A robot that works normally but misses an uncommon human behavior pattern may still be unacceptably risky1.
• Trust comes from how the robot moves. One of the more concrete ideas is that robots should signal intent through intuitive behavior, such as positioning and movement style, rather than relying on lights or buzzers1.

The paper also makes a direct technical recommendation. It favors vision-language-action (VLA) models and transformers for turning high-level semantic intent into physical joint positions2. In other words, the panel is not skeptical of foundation models. It sees them as one part of a larger system that still needs engineering discipline.

The most strategically important claim is about where competitive advantage lies. Panelists suggested that as advanced models become widely available, the edge shifts to execution: systems engineering, disciplined deployment pipelines, and trust built with customers and regulators1. They expect near-term growth in bounded settings, including warehouses, manufacturing, inspection and healthcare support, where labor constraints or safety exposure create strong reasons to automate1.

The foundation-model layer is commoditizing, as the panel predicted

The panel's view that model access is spreading matches how the robot foundation-model market is developing. Market trackers describe three groups competing for this layer. Integrated manufacturers such as Tesla, Figure and 1X train their own models on fleet data. Dedicated labs such as Physical Intelligence, Skild AI and NVIDIA's GR00T project want to supply models that any hardware maker can use15.

NVIDIA's work shows how fast this layer is changing. GR00T N1 was introduced at GTC in March 2025 as an open foundation model for humanoids17. One review says N1.6 shipped at CES 2026 and GR00T 1.7 reached early access at GTC in March 202611. Coverage of open models points to systems such as OpenVLA, Octo and SmolVLA, which are released under permissive licenses, alongside Physical Intelligence's proprietary Pi012. Reporting on the date of GR00T N1's release is inconsistent: one source places it in early 202612, while others date it to March 20251711. That kind of confusion is common when version numbers change this quickly.

If capable VLA models are close to off-the-shelf, the panel's argument follows. The hard part is no longer getting a robot to understand an instruction. It is proving the robot will behave predictably in a crowded aisle on a bad day.

Deployment claims diverge, which is the panel's point

The paper calls for observable, repeatable evidence of performance under real conditions1. Industry coverage shows how uneven that evidence is.

Tesla Optimus is the clearest case. One market tracker reports that more than 1,000 Optimus Gen 3 units have been working in Tesla's Fremont, Austin and Berlin factories since January 2026, with capacity targets of about 100,000 units this year15. A manufacturing-focused outlet repeats the claim of more than a thousand units18. Other trackers are much more skeptical. One classifies Optimus as demo-stage with no verified customer use on record13. Another says Optimus 3 has no official current specifications or price and has not been commercialized11. The two accounts may not strictly contradict each other, since internal use is not the same as external sales. But the gap shows how far public claims can run ahead of independently checkable data.

Figure's record is stronger but also contested. One report says Figure's robots have contributed to the production of more than 30,000 vehicles at BMW's Spartanburg plant, working ten-hour shifts five days a week18. An evidence-tiered tracker counts Figure 02 as customer-confirmed in production but says the newer Figure 03 has not yet reached that bar13.

Agility's Digit is the warehouse benchmark. Multiple outlets agree that Digit has the clearest sustained external deployment record, including GXO and Toyota Canada11. It handles totes and materials in Amazon fulfillment centers and completed a year-long pilot at a Toyota facility18. Agility's RoboFab plant in Salem, Oregon reportedly has capacity for 10,000 units a year13. Still, the numbers vary. One tracker puts Digit at about 250 units in logistics deployments15. Another reports that it moves about 300 boxes an hour, close to 70% of human efficiency, though it can run continuously14. Price estimates range from $50,000–$75,00014 to $100,000–$150,00016.

One guide summarizes the situation in a way that matches the panel's caution. Verified humanoid work in 2026 consists of tote moving, automotive parts handling, inspection and picking. Many systems still combine autonomy with teleoperation or tightly scoped workflows11.

Why warehouses are the proving ground

The panel said embodied AI helps most where machines have to work in spaces designed for people rather than for fixed automation, such as cluttered aisles, varied objects and shared floors with workers1. That describes the warehouse, and it explains why logistics keeps appearing as the first market for humanoids.

The form factor is not settled, though. Skild AI's acquisition of Fetch Robotics, which pairs a wheeled base with manipulation, is described as competing directly with bipedal humanoids for unstructured warehouse tasks19. Galbot's wheeled semi-humanoid is reported in stores in more than 30 cities11. Several industry voices argue that the body is not the real bottleneck. If foundation models can learn from arms, mobile bases and humanoids together, the body becomes a deployment choice rather than a separate AI project20. That also fits the panel's advice to start with the use case, not the technology, and to accept that conventional deterministic automation is sometimes the better answer1.

Standards are the slower-moving piece

The paper's references show its intended direction. It cites ISO work on functional safety for AI systems, AI management systems and the automotive Safety of the Intended Functionality standard356. The panel expects regulators, insurers and customers to demand clearer evidence of readiness, which should speed up shared terminology, evaluation frameworks and lifecycle controls1. Related academic work points the same way. A separate 2026 paper on embodied AI in critical infrastructure argues for bounded autonomy within hybrid human-machine oversight, drawing on the EU AI Act and ISO standards10.

The verdict

The SAE paper does not report a breakthrough. One citation index gives it a low novelty score3, which fits a consensus document. Its value is as a corrective. Humanoid coverage in 2026 is full of unit counts that do not reconcile, version numbers that change every few months, and deployments that range from audited production to demos. Against that background, the panel's insistence on system-level evidence, controlled updates and clear accountability1 is a reasonable standard for judging which claims hold up.

Our reading is that the companies most likely to matter in 2027 are the ones that are already documenting their deployments in detail. Digit's long record with outside customers and Figure 02's hours at BMW count for more than capacity targets. As foundation models become widely available, the panel's closing advice applies: proceed deliberately, but keep moving, because waiting too long is also a competitive risk1.