Safe to deploy, trusted by your operators, defensible to regulators. Orange Box Intel is a physical AI alignment advisory: we check that what your robots and autonomous systems actually do matches what your people intended, before a pilot scales, a deadline arrives, or a near miss becomes an incident.
Nobody budgets for "alignment" in the abstract. Organizations call when a decision is coming and they need to know that the machine will do what they intend. These are the usual triggers.
It worked at one site, with one crew. Now it meets new layouts, new people and new conditions. Will it still pursue the goal you meant?
Risk: goal misgeneralizationEvery demo shows the task completed. You need evidence of how the system behaves when the task, the site rules and the people around it pull in different directions.
Risk: hidden objective misspecificationThe EU Machinery Regulation applies from 20 January 2027, the EU AI Act timeline follows, and UAE permits and sector standards ask for evidence of safe, overseen behavior.
Risk: undocumented intended behaviorThe system "did what it was told", and something still went wrong. You need to know what it was really optimizing for, and what to change before it restarts.
Risk: specification gamingWorkarounds, constant overrides, robots left idle, or the opposite: people who stop watching. Adoption depends on predictable, readable behavior.
Risk: instruction misinterpretationLeadership needs a plain-language view of physical AI risk, liability and the 2027–2028 regulatory timeline before it signs off on capital or a launch.
Risk: alignment risk no one can seeYou buy safety, HSE performance, operator trust and regulatory confidence. With robots driven by learned models, one engineering question underlies all of them: does the system's objective match your intent? That question is alignment. When the answer is no, it shows up in recognizable ways.
In the Dubai Future Foundation's January 2026 report on social robots, workshop participants "stressed the challenge of ensuring value alignment and trustworthiness as robots gain more autonomy", and warned that without "clear liability rules, and robust standards, organisations may avoid or delay adoption". Source: UAE Centre for the Fourth Industrial Revolution, "Social Robots and Society: Global Pathways to Acceptance" (PDF)
Recent robotics research on vision-language-action models reports a consistent "success–safety gap": policies that complete tasks while violating safety constraints along the way. A green checkmark on task completion is not a safety case.
The EU Machinery Regulation brings software and machine-learning-based safety functions into scope, and integrators and importers carry duties of their own. "The model decided" is not a defence.
Standards such as ISO 10218 and IEC 61508 address mechanical and functional safety. They do not tell you whether the model driving the robot interpreted the instruction the way your operators meant it.
The flight recorder is orange, not black. We bring the same discipline to physical AI: what was the machine really optimizing for?
Each engagement answers a question you already have a budget line for, and each targets the alignment failure most likely to undermine that decision. We advise; we do not certify. Our work is built to feed the engineers, lawyers and certification bodies you already use.
When: a pilot is scaling, moving to a new site, or waiting on a go/no-go.
A structured check that the system's objectives, task specifications, guardrails and stop conditions still match what operators and safety owners intend in the conditions it is about to meet.
PREVENTSSpecification gaming · goal misgeneralization
When: you are procuring robots or autonomous systems, or investing in a company that builds them.
The questions to put to vendors, the evidence to require, and an independent read of whether a system's success in the demo reflects the behavior you will need in operation.
PREVENTSReward and objective misspecification hidden behind demo success
When: operators override, work around, over-rely on, or simply avoid the system.
How people actually instruct, supervise, hand over to and anticipate the machine, and where expectations break down. This draws on decades of work on human behavior and nonverbal communication.
PREVENTSInstruction misinterpretation · over-reliance
When: you export to the EU, face a permit or audit, or must show responsible AI in a tender.
One map of what applies to a specific system, and what evidence of intended behavior and human oversight each framework expects, so engineering, legal and procurement work from the same page.
PREVENTSUnstated assumptions about intended behavior that stall conformity assessment
When: a new task, instruction interface or AI model is about to go live.
An adversarial review of task specifications, objectives and natural-language instructions. The question is how this goal could be satisfied in a way nobody wanted. We stress-test intent before the robot does.
PREVENTSSpecification gaming · instruction misinterpretation
When: a board, C-suite or investment committee needs to decide.
A private session on physical AI risk in plain language: what can go wrong between the task you specify and the behavior you get, where liability sits, and what the 2027–2028 timeline means for your programs.
PREVENTSAlignment risk going unseen by the people who sign off
Incidents and stalled deployments rarely trace back to one broken part. More often they trace back to a gap between what people intended and what the system was told to optimize. The Orange Box method traces that chain end to end.
INTENT → SPECIFICATION → BEHAVIOR → PEOPLE
Capture what operators, safety owners and leadership actually mean the system to do, including the unwritten rules.
Map that intent against objectives, rewards, task specs, prompts, guardrails and stop conditions. Find what was never written down.
Review test evidence and operating scenarios for shortcuts, ambiguity and "success-but-unsafe" outcomes.
Examine how humans instruct, supervise and override the machine, and where their expectations break down.
Robots and autonomous systems are now running in UAE plants, streets, buildings and airports. Dubai's Robotics and Automation Program targets 200,000 robots by 2032. Every new deployment adds another place where behavior has to match intent.
Source: UAE Government portal, Dubai Robotics and Automation Program
In the UAE: ADNOC has deployed Taurob's heavy-duty inspector robot for autonomous inspections at its Taweelah Gas Compression Plant, and is co-developing an "operator" robot to lift equipment and turn valves, due to be operational by the end of 2026.
Abu Dhabi Media Office, 21 May 2026
The alignment question: when a robot moves from looking to acting, does "complete the round" still mean "complete it the way your HSE rules intend"?
In the UAE: Dubai's RTA began commercial autonomous taxi operations in Umm Suqeim and Jumeirah in March 2026, starting with 100 vehicles. In Abu Dhabi, WeRide's driverless service on Yas Island is ITC-certified.
RTA, 30 Mar 2026 · Abu Dhabi Mobility
The alignment question: how does the driving policy handle situations that local passengers, operators and regulators treat as obvious?
In the UAE: Abu Dhabi's ITC is supervising a pilot of autonomous trucks on dedicated KEZAD routes with AD Ports Group. Talabat is testing delivery robots that ride elevators in a Dubai office building. Emirates expects to launch check-in and baggage robots in the first quarter of 2027.
WAM, 27 Feb 2026 · Khaleej Times, Sep 2026 · Gulf News, 22 Sep 2026
The alignment question: at the handoffs between people and machines, does everyone read the robot's intent the same way?
In the UAE: Dubai's Robotics and Automation Program lists developing an act for the use of robots among its initiatives. Dubai Police deployed its autonomous patrol DPR 02 at Global Village in October 2025.
UAE Government portal · WAM, 14 Oct 2025
The alignment question: what behavioral evidence should permits, sandboxes and public procurement ask for, beyond hardware safety?
In the UAE: Presight (a G42 company) and Shorooq backed industrial robotics company Maven Robotics in a $100M Series A in September 2026. Gulf capital is moving into physical AI.
The alignment question: beyond uptime and reliability figures, can the team show what its robots optimize for, and how it would know if that drifted?
In the UAE: Abu Dhabi's Department of Health Responsible AI Standard (2025) applies across the emirate's healthcare ecosystem, including AI sourced from vendors, and sets out human oversight, validation and traceability requirements. The revised ISO 13482 for service robots is at final-draft stage.
DoH Responsible AI Standard (PDF) · ISO 13482
The alignment question: will the oversight design hold up with patients, visitors and vulnerable users, not just trained staff?
Accredited certification bodies test hardware hazards and functional safety, against standards such as ISO 10218 and IEC 61508, and certify AI management systems under ISO/IEC 42001. That work is essential, and we do not replace it.
Alignment asks a separate question that sits upstream of certification: whether a learned system's objectives and instructions produce the behavior your people intend. Our gap registers and red-team scenarios are written to feed your hazard analysis, safety case and AI management system, so you arrive at certification better prepared.
If you sell robots into the EU, buy from vendors who do, or operate under UAE sector rules, these are the dates and frameworks your program will be judged against. Status as of November 2026. This is general information, not legal advice.
Dubai Law No. 9 of 2023 regulates the operation of autonomous vehicles. Abu Dhabi's Department of Health Responsible AI Standard (2025) sets requirements for human oversight, validation and traceability across the emirate's healthcare ecosystem.
Replaces the Machinery Directive 2006/42/EC. Software can be a safety component, and machinery with self-evolving, machine-learning-based safety functions falls into categories that require third-party conformity assessment.
Following the AI Omnibus, obligations for high-risk use cases such as critical infrastructure apply from this date: risk management, logging, documentation, human oversight, robustness and cybersecurity.
Rules for high-risk AI in regulated products apply from this date. Under the Omnibus, AI in machinery is addressed through the Machinery Regulation rather than directly through the AI Act.
ISO 25785-1, for dynamically stable industrial mobile robots including legged humanoids, is at committee-draft stage. The revision of ISO 13482 for service robots is at final-draft stage. Neither is published yet.
ISO 10218-1/-2:2025ISO/TS 15066ISO 13482UL 3300ISO 25785-1 (draft)
IEC 61508ISO 21448 (SOTIF)
ISO/IEC 42001EU AI ActUAE AI Charter
Dubai Law 9/2023 (AVs)DoH Responsible AI StandardDubai R&A ProgramDIFC Regulation 10UAE AI Strategy 2031
Orange Box Intel is a new advisory practice founded by Jesús Enrique Rosas, founder of Knesix (KNESIX FZ LLC, UAE).
Known publicly as The Body Language Guy, Jesús has spent decades studying human behavior, nonverbal communication, persuasion and negotiation, and training professionals to read the gap between what people say and what they mean.
Orange Box Intel applies that lens to physical AI. Aligning robots with human intent is, at its core, a communication problem: instructions that are ambiguous, expectations that are never written down, and signals people send without words. The practice brings that behavioral perspective together with the safety standards and regulation that govern machines.
Alignment is not only an engineering property. It is a relationship between a machine and the people who rely on it.
Short answers to the questions we hear most. For anything specific to your systems, request a briefing.
Because it sits underneath what you already buy: safety, HSE performance, operator trust and regulatory readiness. With robots driven by learned models, those outcomes depend on whether the system's objectives match what your people intend. Alignment is the name for that match, and it is what we review.
It is the work of making sure robots and other embodied AI systems pursue the goals people actually intend, within the constraints people actually expect. It sits between AI safety research, functional safety engineering and human factors.
A goal or task discrepancy is a gap between the objective a system was given and the outcome its owners wanted. In software, that produces a wrong answer. In a robot, it produces a wrong movement: a shortcut through a restricted area, excessive force, or a task "completed" in a way that leaves something unsafe. Known forms include specification gaming, reward misspecification, instruction misinterpretation and goal misgeneralization.
Background: Google DeepMind on specification gaming
As a complement. Certification bodies test functional and hardware safety and certify AI management systems such as ISO/IEC 42001. We work upstream on whether behavior matches intent, and we write our outputs to feed your hazard analysis and safety case. We are not a certification or notified body and do not issue certificates.
Regulation (EU) 2023/1230 applies from 20 January 2027 and replaces the Machinery Directive. It recognizes software as a possible safety component, and it places machinery whose safety functions rely on self-evolving, machine-learning behavior into categories that require third-party conformity assessment. If you sell robots into the EU, or buy from vendors who do, it applies to you directly.
Yes. Under the AI Omnibus, high-risk rules for Annex III use cases apply from 2 December 2027, and rules for high-risk AI embedded in regulated products (Annex I) apply from 2 August 2028. AI in machinery is now addressed through the Machinery Regulation.
Functional safety testing asks whether safety functions work when needed. An alignment review asks a different question: whether the system's objectives and instructions lead it to behave the way your people intend in the first place. Robotics research on vision-language-action models shows that high task success can coexist with unsafe execution. The two reviews complement each other.
Not a published one. ISO 25785-1, covering dynamically stable industrial mobile robots such as legged humanoids, is still a draft. Until it is published, organizations rely on ISO 10218:2025, related standards and their own risk assessment, so a clear view of intended behavior matters even more.
ISO/IEC 42001 is a management-system standard for organizations that develop or use AI, regardless of form factor, so the AI components of a robotic system can fall under an organization's AI management system. In the UAE it is voluntary, but it increasingly appears in procurement requirements.
Most begin with an executive briefing tied to a specific trigger, such as a scale-up, a vendor selection or a deadline. That is followed by a scoped review of one system, site or procurement. Scope, timing and deliverables are agreed up front.
A pilot scaling, a vendor to choose, a deadline, an incident, or a board question. Tell us the system and the decision, and we will reply by email to arrange a conversation.
Helpful to include: your organization, the robotic or AI system involved, where it operates, the decision you face, and your timeline.