Cutting-Edge Safety Technology: Vision-Language Models, Physical AI, Robotics and EHS Copilots

This chapter looks at the technologies that sit just past today's mainstream safety tools: vision-language models that can describe and search video, physical AI systems such as humanoid robots and driverless terminal tractors, digital twins, exoskeletons, drones, generative AI copilots inside EHS software, synthetic training data and the edge computers that run all of it. Most are real and in early commercial use, but the evidence that they reduce injuries is thin, so buyers should treat them as pilots with clear success criteria rather than proven controls.

For each technology, the chapter covers what it does, where it has been deployed publicly, what the independent evidence says and what a safety team should ask before committing budget. Vendor statements are attributed to the vendor, and where no independent evaluation exists, we say so.

What is "cutting edge" in safety technology in 2026?

Chapters 4 to 7 covered tools that many large employers already run in production: AI video analytics on existing CCTV, wearables, proximity warning systems and EHS software. The technologies in this chapter differ in three ways.

First, they rely on newer kinds of AI. Earlier safety analytics used narrow models trained to spot one thing, such as a missing hard hat or a person inside a marked zone. The newer systems use foundation models trained on very large, general datasets that can be prompted in natural language to handle tasks they were not specifically trained for.

Second, several of them act in the physical world. A detector that raises an alert leaves the decision to a person. A robot or a driverless tractor makes and executes decisions itself, which brings it under machinery safety law and standards.

Third, the evidence base is younger. For AI video analytics there are now multi-year deployments and some published before-and-after data, although much of it comes from vendors. For humanoids, vision-language models and copilots, public data is mostly from pilots, press releases and lab studies.

How are vision-language models changing safety video analytics?

A vision-language model (VLM) takes images or video plus text as input and produces text as output. You can show it a frame from a loading dock and ask "Is anyone standing behind the reversing truck?" and it will answer in a sentence. Large general-purpose models from the major AI labs have had this ability since 2023, and smaller open models that can run on site have followed.

What can a VLM do that a trained detector cannot?

A conventional safety detector is trained on thousands of labeled examples of one event type. It is fast and measurable, but it only knows what it was trained on. Adding a new rule, such as "flag anyone using a phone while operating a reach truck," usually means collecting and labeling new data.

A VLM can be asked about things nobody labeled in advance. That opens up three practical uses.

The first is natural-language search. Instead of scrubbing hours of footage after an incident, an investigator can type "forklift passing within a meter of a pedestrian near door 4 yesterday afternoon" and get candidate clips. NVIDIA's public reference design for this, the AI Blueprint for video search and summarization (VSS), combines VLMs, large language models and retrieval so that users can search live or recorded video in plain language, summarize long recordings and ask questions about what happened ^[1]. NVIDIA describes uses across industries including safety ^[2], and software companies have built products on top of it; Superb AI, for example, has announced an integration of the blueprint into its video monitoring platform ^[27].

The second is alert verification. A fast, cheap detector raises a candidate event, and a VLM looks at the clip to confirm or reject it before a person is notified. NVIDIA lists "real-time verified alerts" as one of the VSS blueprint's functions ^[1]. The goal is fewer false positives reaching supervisors, which addresses the alert fatigue problem discussed in Chapter 4.

The third is narrative reporting. A VLM can turn an event clip into a short written description ("A pedestrian walked through the marked forklift aisle while a counterbalance truck was reversing; no contact occurred"), which can pre-fill a near-miss report.

Safety analytics vendors have signaled the same direction. When Intenseye announced its $64 million Series B in February 2024, the company said it planned to bring large language models to safety teams and to invest in a generative AI video anonymization feature ^[3].

How accurate are VLMs at spotting hazards?

Independent studies are still small, but they point the same way: general-purpose VLMs can recognize many hazards, miss a meaningful share and need tuning to perform well.

A 2026 study in the journal Information tested a large language model with vision on construction site hazard identification against expert judgment. The authors reported that the model identified hazardous situations about 69.6 percent of the time, missed about 30.4 percent (false negatives) and agreed with the expert consensus in about 68.6 percent of its predictions ^[4]. A miss rate near one in three would be unacceptable for any control that people rely on.

Fine-tuning helps. Researchers who built ChatCH, a construction hazard model fine-tuned from the open Qwen2-VL-7B model, reported a precision of 89.4 percent, well above the same base model without fine-tuning ^[5]. That result came from the authors' own test set, and precision alone does not tell you how many hazards were missed.

The practical reading for buyers is that a VLM is a good assistant for search, triage and description, and a poor sole detector for life-critical events. Ask any vendor that uses VLMs to separate the two roles and to provide precision and recall figures measured on footage from sites like yours.

What new risks do VLMs bring?

VLMs produce fluent text even when they are wrong. A wrong answer from a detector is a box in the wrong place; a wrong answer from a VLM is a confident sentence that may end up in an incident record. They are also sensitive to how a question is worded, so two supervisors asking slightly different questions can get different answers about the same clip.

There are privacy questions as well. A system that can answer arbitrary questions about video can be used to answer questions about individual behavior that have nothing to do with safety. Chapter 13 covers data protection impact assessments and purpose limitation; a VLM deployment needs a written list of permitted query types, logging of who asked what, and a review of whether prompts could identify individuals.

What is physical AI and why does it matter for safety?

"Physical AI" is the industry term for AI systems that perceive and act in the physical world: robots, autonomous vehicles and the models that control them. NVIDIA has used the term heavily since 2024, and its product launches give a reasonable picture of where the technology stack is heading.

At CES in January 2025, NVIDIA announced Cosmos, a family of "world foundation models" designed to generate physics-aware synthetic video and sensor data for training and testing robots and autonomous vehicles. NVIDIA said Cosmos would be released under an open model license and named early adopters including Agility, Figure AI, 1X, Skild AI, Uber and Waabi ^[6]. In August 2025 it made the Jetson AGX Thor developer kit available at $3,499, a robot computer that NVIDIA says delivers up to 2,070 FP4 teraflops of AI compute within a 130-watt power envelope ^[7].

For safety managers, physical AI matters in two ways. It can take people out of hazardous tasks, which sits at the elimination and engineering levels of the hierarchy of controls. It also introduces machines that move less predictably than fixed automation, which creates new collision, crushing and entanglement risks that existing risk assessments may not cover.

How are digital twins used for safety?

A digital twin is a software model of a physical facility, process or asset that is kept updated with real data. The twins that are useful for safety fall into two groups.

Design and planning twins let engineers simulate a warehouse, terminal or production line before building or changing it. They can test traffic routes, robot paths, pedestrian crossings and sightlines, then adjust the layout before anyone is exposed. In January 2025 NVIDIA released "Mega," an Omniverse blueprint for building large-scale digital twins of industrial robot fleets ^[8]. KION Group, Accenture and NVIDIA said they would use Mega so that warehouse operators can design efficient and safe configurations, including the number of robots, workers and pieces of equipment, without interrupting live operations ^[9].

Operational twins take live feeds from cameras, sensors, vehicle telematics and wearables and show the current state of a site. Combined with the analytics in Chapters 4 to 6, they can show where near misses cluster, how traffic flows change by shift and where congestion builds before incidents.

Two cautions apply. A twin is only as accurate as the data behind it; an outdated layout or a missing temporary structure makes its conclusions wrong. And simulated behavior of people is still crude. Workers take shortcuts, stand in odd places and react to pressure in ways that most simulations do not model. Use twins to rule out bad designs early, and validate the chosen design with real observation after go-live.

Will robots and humanoids make workplaces safer?

Robots have reduced exposure to some hazards for decades, from welding cells to automated storage and retrieval systems. The newer developments are mobile robots that share space with people and humanoids designed to work in facilities built for humans.

What do current robot safety standards require?

The core international standard for industrial robot safety, ISO 10218, received its first major revision since 2011 in February 2025. According to the Association for Advancing Automation (A3), the 2025 editions make functional safety requirements more explicit, fold in collaborative robot guidance that previously sat in the separate technical specification ISO/TS 15066, and add guidance on manual loading and unloading and on end-effectors ^[10]. Part 1 covers robot manufacturers and Part 2 covers integrators who build robot applications and cells.

In the EU, the new Machinery Regulation (EU) 2023/1230 applies from January 20, 2027. It explicitly covers safety components with self-evolving behavior that use machine learning, and adds requirements on cybersecurity and autonomous machinery ^[26]. Any robot or autonomous vehicle bought for EU sites from 2027 onward will need to meet it.

What has been deployed in practice?

Ground inspection robots are the most established category. Boston Dynamics' Spot quadruped has been used for routine inspection rounds in industrial plants. In a Boston Dynamics case study, Cargill's Amsterdam Multiseed facility, which had relied on technicians walking routes with handheld thermal and acoustic detectors, said Spot found an air leak on a centrifuge within minutes of being powered up ^[13]. In June 2026, MFE Inspection Solutions announced a package that mounts a Blackline Safety connected gas detector on Spot and streams readings, alerts and location data to remote teams ^[29]. Deployments like these are meant to cut the number of times people walk into areas with gas, heat or moving equipment.

Humanoids are newer. GXO Logistics says it signed the industry's first multi-year robots-as-a-service agreement for humanoid robots with Agility Robotics, deploying Digit at a facility operated for the apparel brand Spanx in Georgia after a 2023 pilot ^[11]. Digit moves totes between collaborative robots and conveyors. At BMW Group's Spartanburg, South Carolina plant, Figure AI says its Figure 02 humanoid completed an 11-month deployment, running 10-hour shifts on weekdays, loading more than 90,000 sheet metal parts and contributing to the production of more than 30,000 BMW X3 vehicles ^[12]. Both are vendor accounts.

What safety questions do humanoids raise?

Humanoids are tall, heavy, battery-powered machines that walk. A fall or an unexpected arm movement near a person is a different risk from a fixed robot behind a guard. Few published incident statistics exist for humanoids, so a buyer has no industry baseline. Questions worth putting to any humanoid supplier include:

• Which standards has the robot been assessed against, and by whom?
• What happens when it loses balance, power or network connection near a person?
• What speed and force limits apply when people are within reach, and how are they enforced?
• Is it deployed in a zone segregated from pedestrians, or does it share aisles?
• How will incidents and near misses involving the robot be recorded and reported to you?

The safest early uses keep humanoids in segregated cells or low-traffic areas, doing tasks where a stop does not create a new hazard.

Do exoskeletons reduce musculoskeletal injuries?

Exoskeletons are wearable devices that support the body during lifting, bending or overhead work. Passive designs use springs or elastic elements; active designs use motors and batteries. Musculoskeletal disorders remain one of the largest categories of workplace injury cost, so interest is high.

The evidence is mixed. A systematic scoping review published in 2025 found that exoskeletons were used most in manufacturing and perioperative healthcare, that they reduced muscle load during repetitive or static tasks, and that adoption was limited by discomfort and fit problems ^[15]. Most studies in the review were short-term and lab-based. NIOSH researchers reported in October 2025 on shoulder-assist exoskeletons tested during block-laying on simulated mast climbers, a construction work platform; they found the devices gave minimal and inconsistent reductions in shoulder muscle strain for those tasks ^[14].

Powered devices are getting stronger. German Bionic launched its Apogee ULTRA powered exoskeleton in January 2025 and says it provides dynamic lifting support of up to 36 kilograms (80 pounds) ^[17]. The company also claims that customers using its exoskeletons have seen lower sick leave; that figure is a vendor claim and has not been independently verified.

Standards are still forming. ASTM Committee F48 on Exoskeletons and Exosuits was formed in 2017 and develops voluntary standards covering safety, performance, ergonomics and terminology for industrial and other uses ^[16].

An exoskeleton is a form of PPE in practice, so it sits low on the hierarchy of controls. Before buying, check whether the task can be redesigned with lift assists, conveyors or better layout. If you do pilot exoskeletons, involve the workers who will wear them, test over weeks rather than hours, and watch for load transferred to other joints, heat stress, balance problems and restricted movement in emergencies.

Are autonomous yard and port vehicles safer than human drivers?

Yards and container terminals are high-risk places for vehicle-pedestrian contact. Trucks and terminal tractors reverse frequently, sightlines are poor and drivers get in and out of cabs to couple trailers. Automating these moves can remove people from the most dangerous part of the job, which is being on foot among moving vehicles.

Two companies show where the technology stood in 2025 and 2026.

Outrider, a US company focused on distribution yards, raised a $62 million Series D in late 2024 led by Koch Disruptive Technologies and NEA, with participation from NVIDIA's venture arm and Prologis Ventures, bringing its total raised to more than $250 million ^[18]. The company said its technology had been validated by more than 100,000 autonomous trailer moves and that it planned to scale with Fortune 500 customers in 2025 ^[18]. Outrider says its system automates tasks that put yard drivers on foot, such as connecting air lines between tractor and trailer.

FERNRIDE, based in Germany, said in July 2025 that its autonomous terminal tractor platform had received TÜV SÜD certification of its safety concept and system design under the EU Machinery Directive, and approval from the Estonian Transport Administration, allowing the start of driverless operations at the HHLA TK Estonia terminal near Tallinn with three tractors ^[19]. FERNRIDE says it was the first company to receive such certification for an autonomous terminal tractor.

The main safety risk during adoption is mixed traffic. When autonomous vehicles share space with manned trucks and pedestrians, everyone's expectations change. Ask vendors how their vehicles detect and respond to people on foot, what the minimum stopping distance is at operating speed, how remote operators intervene, and what happens if communications fail. Update traffic management plans and site rules before the first driverless move.

How are drones and remote inspection used for safety?

Drones reduce work at height and confined-space entry by sending a camera instead of a person. Common uses include roof and stack inspections, stockpile surveys, flare stack checks, bridge and tower inspections and post-incident surveys of damaged structures.

The newer development is docked drones that launch automatically from a charging station on site and can be flown by a remote pilot. Skydio says the FAA has granted waivers to Dominion Energy and the New York Power Authority for fully remote beyond-visual-line-of-sight (BVLOS) operations using its docks, with no flight crew on site ^[21].

US regulation is the main constraint. Routine BVLOS flights currently need a Part 107 waiver or another case-by-case FAA approval ^[28]. The FAA published a proposed rule for BVLOS operations, Part 108, on August 7, 2025, covering drones up to 1,320 pounds and requiring detect-and-avoid capability and Remote ID ^[20]. According to a status guide from the drone training organization USI, the final rule went to the White House Office of Information and Regulatory Affairs for review in July 2026 and had not been published as of mid-September 2026 ^[28]. Plans that depend on routine remote drone operations should assume waivers will be needed until the final rule takes effect.

Drones also create hazards of their own: falling aircraft, propeller strikes during handling and battery fires. They capture images of workers and neighbors, which brings the privacy considerations in Chapter 13 into play.

What can generative AI copilots in EHS software actually do?

Since 2024, most major EHS software vendors have added generative AI features. The product names differ, but the functions are similar.

In practice, copilots help with tasks such as:

• Turning a voice note or free-text description into a structured incident or near-miss report.
• Suggesting incident categories, body parts and severity codes.
• Flagging reports that may describe a potential serious injury or fatality (SIF) precursor.
• Summarizing trends across hundreds of reports.
• Drafting job hazard analyses, toolbox talks and procedure updates from existing documents.
• Answering questions about internal procedures or regulations, with links to source documents.

Copilots can save time, especially for frontline staff who find forms slow. They can also get things wrong: a copilot can classify an event wrongly, invent a plausible root cause or summarize a regulation inaccurately. Errors in recordkeeping can affect OSHA logs, insurer reporting and legal cases.

A sensible governance approach treats copilot output as a draft. The person who submits a report or approves an investigation remains responsible for its content. Ask vendors which models they use, where data is processed, whether your data trains shared models, how outputs are logged and whether the system shows its sources. Cority, for example, has positioned Cortex AI around a central control center for governing AI use ^[23]; whatever the vendor's claim, test the controls yourself.

Can synthetic data solve the rare-event problem?

Serious incidents are rare, which is good for workers and bad for model training. A detector for a person caught between a forklift mast and a rack needs examples, and real footage of that event is scarce and sensitive. Synthetic data, meaning images or video generated by simulation or generative models, is one way to fill the gap.

NVIDIA positions Cosmos as a way to generate large amounts of photoreal, physics-based synthetic data to train and evaluate models ^[6]. Simulation tools built on game engines and 3D platforms have been used for this purpose for years.

Research shows that synthetic data helps most when mixed with real data. A 2026 study in the journal Buildings tested PPE detection models including YOLOv11s, Faster R-CNN and RT-DETR across 11 different mixes of real and synthetic training images, evaluated on a fixed set of 400 real images ^[25]. The design reflects a common finding in this research: models trained only on synthetic images tend to perform worse on real footage, so the practical question is how much synthetic data to add to a real dataset.

For buyers, synthetic data is mostly a vendor engineering choice. It becomes your concern when a vendor claims a model can detect an event it has never seen on a real site. Ask how the model was validated on real footage and with what results.

What edge hardware will run these systems?

Edge AI means running models on hardware at or near the camera or machine, instead of sending everything to a distant cloud. It reduces latency, network cost and the amount of video leaving the site, which helps with privacy.

Hardware has improved quickly. NVIDIA's Jetson AGX Thor, available from August 2025, is aimed at robots but shows the trend: NVIDIA says it delivers up to 7.5 times the AI compute of its predecessor, Jetson Orin, with 3.5 times better energy efficiency, and enough memory to run generative models on device ^[7]. Outlook: as edge compute grows, more VLM-based verification and search is likely to run on site rather than in the cloud.

For safety buyers, the practical questions are about lifecycle. Edge devices need patching, monitoring, replacement and secure configuration. Ask who owns the hardware, how long the vendor supports it, how firmware updates are delivered and what happens to the analytics if a device fails.

How should a safety team evaluate emerging technology?

The technologies in this chapter share a pattern: real capability, strong vendor narratives and limited independent outcome data. The buying and piloting discipline in Chapter 14 applies with extra force.

Start from the risk

Pick the hazard first. If your serious injury data shows vehicle-pedestrian interactions in the yard, autonomous yard trucks and better traffic separation are relevant. If your data shows overhead work strain, exoskeletons and task redesign are candidates. Technology bought without a target risk tends to end up as a demonstration that controls nothing.

Place it on the hierarchy of controls

Robots and autonomous vehicles that remove people from a hazard are engineering controls or elimination. VLMs, copilots and dashboards support administrative controls. Exoskeletons behave like PPE. Prefer options higher on the hierarchy, and do not let an analytics tool substitute for a physical control.

Define success before the pilot

Write down what the pilot must show. For a VLM verification layer, that might be a target reduction in false alerts with no increase in missed events, measured against human review. For an autonomous yard vehicle, it might be a defined number of moves with no unplanned pedestrian interactions inside a set distance. For a copilot, it might be report completion time and the error rate found in a sample audit.

Check the legal and standards basis

Confirm which standards a machine has been assessed against (ISO 10218 for industrial robots, the EU Machinery Directive or the new Machinery Regulation from 2027, sector rules for vehicles and aircraft). For AI that processes video or worker data, complete the privacy and labor law checks in Chapter 13, including the EU AI Act where it applies. After the Digital Omnibus amendment entered into force on July 27, 2026, AI Act obligations for stand-alone Annex III high-risk systems apply from December 2, 2027, and for AI embedded in Annex I products such as machinery from August 2, 2028 ^[30]. Chapter 17 covers these dates in more detail.

Involve workers early

Workers who share space with robots, wear exoskeletons or are recorded by cameras will judge whether the technology is trustworthy. Early involvement reduces resistance and surfaces practical problems that vendors miss.

Plan for failure modes

Every system here can fail. Ask what happens when a model is wrong, a robot stops, a drone loses its link or a copilot gives bad advice. A good deployment fails safe and makes failure visible.

Questions to ask vendors of emerging safety technology

Summary

Vision-language models, physical AI, digital twins, robots, exoskeletons, autonomous yard vehicles, drones, EHS copilots, synthetic data and faster edge hardware are all moving from research into commercial use. Each has a plausible route to fewer injuries, usually by removing people from hazards or by helping safety teams see and act on risk faster.

The evidence is still early. Published studies show general-purpose VLMs missing a substantial share of construction hazards, exoskeleton benefits vary by task, and the best-known humanoid and autonomous vehicle deployments are described mainly by the vendors themselves. Standards and law are catching up, with ISO 10218 revised in 2025, the EU Machinery Regulation applying from January 2027 and the FAA's BVLOS rule still pending in late 2026.

Safety teams that do well with these technologies will start from a specific risk, place each tool on the hierarchy of controls, define success before a pilot, check the standards basis, involve workers and plan for failure. Chapter 17 looks at how regulation, insurance, markets and the workforce are likely to shape adoption through 2030.