AI systems which work autonomously are slowly starting to leave the virtual world and enter the physical one such as factories, logistics operations, and the general public. This process is raising the question whether present laws covering AI address embodied AI systems.
Up until now, most laws concerning the regulation of AI technology have been geared towards mitigating threats in the virtual world, which includes issues like discrimination, disinformation, and content-related concerns. There are dangers associated with embodied AI, especially when failure leads to physical consequences to infrastructure, personal property, and even human beings.
Version 1.5 of Singapore’s Infocomm Media Development Authority Model AI Governance Framework for Agentic AI was released on May 20th. The framework describes how an organization should guide its AI agents that plan, make decisions, and carry out those decisions to achieve desired outcomes for users through multiple steps.
Agents, according to the framework, are able to interact with tools, external programs, and other agents, including any system that performs actions such as database updates, file creation, device control, and transactions.
Artificial Intelligence is applied to the physical world
During a recent AI summit in Singapore, the conversation about robotics and embodied AI was centered around issues related to operational safety, which are typically seen in areas such as aviation, industry, and critical infrastructure, rather than traditional software regulation.
Moreover, there was discussion about whether autonomous systems could reliably run in unpredictable and changing real-life conditions over long stretches of time.
According to Dr. Ya-Qin Zhang, founding dean of the Institute for AI Industry Research at Tsinghua University, embodied AI poses additional risks beyond those of autonomous software, as any malfunction would potentially impact systems of transportation, drones, logistics, and infrastructure.
“Any risk in the digital world becomes a risk in the physical world and the risk in the physical world becomes a physical risk,” Zhang told MLex on the sideline of the summit.
In particular, Zhang suggested that vehicles, drones, power grids, and other critical infrastructures might become vulnerable once AI is increasingly integrated with physical activities.
The issue of reliability, operational monitoring, and post-deployment monitoring were brought up during speakers’ remarks regarding governance issues. The Summit suggested that for governance based on deployment, simulation, telemetry, and testing will be needed, rather than only one time certification.
The IMDA’s guidelines recommend phased roll-outs, monitoring, and additional tests even after deployment. The IMDA asserts that the agent is dynamic and reacts to its environment, hence some risks cannot be determined prior to its deployment.
Also Read : NVIDIA’s Vera Chip Is a $200 Billion Bet Jensen Huang Wants the World to Notice
It then becomes an operational matter
The Grab company, which has tested self-driving cars and robots for deliveries in Singapore’s Punggol area, said that governance in deployment relied significantly on simulation, testing, and monitoring.
“I mean we simulate a lot. We test a lot on both closed circuits and open circuits to ensure that our robots are reliable,” said Suthen Thomas Paradatheth, Chief Technology Officer of Grab, at one of the summit discussions.
“And before we go from a few robots to hundreds of robots, we ensure that we get it right through simulation and by testing just a few robots first,” he continued.
At the same time, Grab noted the use of monitoring technologies to assess robot performance and uncover potential problems arising from the robot’s usage in the real world.
“This is because there are quite a number of issues which could pop up,” Paradatheth explained.
According to the IMDA framework, agentic use cases should be evaluated in terms of their data access, access to external systems, level of autonomy, and complexity of tasks involved.
Additionally, it also emphasizes that there should be restrictions on the access of agents to the systems. Organizations must implement an off mechanism in case their agents fail.
Accountability is spread over a wider range of individuals
According to MLex, an embodied AI system could consist of more than one entity in its development, production, and implementation processes. These entities include the AI developer, robotics manufacturer, semiconductors supplier, and the infrastructure provider.
MLex also adds that the assignment of responsibility becomes more difficult once systems keep learning and adapting after deployment using software updates, telemetry, and other information about operations.
As per IMDA, entities and individuals are responsible for the agent’s activities irrespective of its autonomy. It states that an agentic AI value chain needs accountability at each stage, including model and platform provider, implementer, tooling provider, and user.
In Applied Materials’ view, deployment of robotics is not only associated with semiconductor technology but also related to systems integration aspects. Om Nalamasu, CTO of the firm, stated that robots’ systems will need improvements in sensors, energy consumption, packaging, and computing architecture.
According to Nalamasu, robots would need a dedicated design process for specific industries and ecosystems and not a generic approach suitable everywhere.
According to Zhao Yuli, the Chief Strategy Officer of Chinese robotics startup company Galbot, Beijing is focusing on deploying robots in scale and industry applications through government-sponsored testing beds, industry collaborations, and long-term funding initiatives.
The company Galbot is deploying humanoid robotics in China in sectors like retail, warehouses, and pharmaceutical applications. This includes fully autonomous stores which can run for 24×7. According to Zhao, semi-structured industrial application settings are expected to be one of the earliest areas of commercialization due to their controllable environment.
On the other hand, Japanese researchers have focused their efforts towards standards development, robotics data sets, and safety-related governance. Professor Yutaka Matsuo from the University of Tokyo’s Graduate School of Engineering mentioned an “AI Association” project wherein 100,000 hours of robotics data was expected to be collected to create robotics foundation models.
Furthermore, Matsuo also highlighted the activities being carried out by the AI Safety Institute in Japan along with the Hiroshima AI Process, involving Singapore and other Asian countries.
Also Read : Elon Musk and Mark Zuckerberg Reportedly Convinced Donald Trump to Scrap an AI Executive Order
The country of Singapore provides rules for agents
There are four categories in Singapore’s framework that specify agency of AI from the governance perspective. Those include initial risk evaluation, human oversight and accountability, technological controls, and the end-user responsibilities.
According to the document, human oversight has to be rethought for agentic AI since continuous inspection of every step in the workflow is unrealistic at scale. Human approval is suggested at crucial moments like performing critical operations, irrevocable actions, or any other deviations from usual performance.
IMDA also highlights potential problems associated with human supervision of highly capable agents like automation bias or alert fatigue. Auditing oversight by evaluating the frequency of human intervention and its duration is proposed. Moreover, the agency suggests automated real-time monitoring of agent activities to detect outliers.
As for user guidelines, the framework states that users have to be informed about the range of agent actions, available data resources, and responsibilities left for users to perform. The agency also stresses the importance of employee training in the human-agent communication, supervisory activities, and professional analysis of agent output.
AI is tested by firms in regulated processes
According to Paul Uren, JPMorgan’s head of investment banking for Asia-Pacific, the bank was rolling out AI technology across its global investment banking operations, Reuters reported. The bank explained that this was meant to allow bankers to access more information and synthesize it in their internal systems. It was also helping to create content and engage clients.
According to Jamie Dimon, CEO of JPMorgan, the bank would increase hiring AI experts instead of traditional bankers. Reuters reported that globally, banks are ramping up AI investments, retooling their workforce, and shifting job responsibilities.
The bank is also one of a few organizations allowed to test Anthropic’s Mythos cybersecurity tool under an initiative called Project Glasswing. As per Anthropic, Mythos can detect any old vulnerabilities within browsers, infrastructures, and applications.
Reuters reported that other banks such as Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley were accessing, or testing, Mythos according to sources and executives at the respective organizations.
IMDA’s report also has a case study about the source-of-wealth analysis conducted by OCBC Bank of Singapore. The tool analyzes income documents and prepares a source-of-wealth memorandum. It doesn’t make decisions on credit, onboarding, or risks independently.
In such a case, the workflow is limited to task-level autonomy and runs solely on the basis of pre-defined workflows. Human intervention is essential at crucial points of decision-making, with validation still left to the reviewers.
Robots enter industrial applications
According to a recent survey by Nikkei Research carried out for Reuters, one in three firms in Japan has either deployed AI robots or is seriously thinking about it. Conducted between May 1 and May 15, the survey polled 492 firms anonymously.
Some 4% of respondents said that they already have AI robots, 5% plan to introduce them into their operations, and another 25% are considering doing so. Some 66%, on the other hand, have no plans to deploy AI robots whatsoever.
As far as the respondents’ industries go, the companies operating in transportation equipment manufacturing have the highest percentage of those who have either already introduced the technology or are going to do so – 80%. At the same time, 94% of those from the wholesale sector do not intend to implement this solution at all.
When asked how they use or plan to use AI robots, some 71% picked manufacturing as the application area, followed by dangerous tasks (19%), and customer services (11%).
The introduction of AI robots will be useful to the Japanese authorities as an instrument to cope with the persistent labour shortage problem and retain a leading position in robotics development.
Retail agents go further than searching for
The retail giant Walmart has presented their roadmap where agentic AI can be utilized within customer, associate, supplier, and developer flows.
Specifically, Walmart has unveiled plans for their next generation of four different AI-enabled ‘super agents.’ These will be for customers, store associates, suppliers, and software developers. According to Walmart, the AI-driven super agents will serve as their primary entry point to engage customers in interaction via artificial intelligence.
In terms of tools, Walmart has already deployed Sparky within its application as a generative AI assistant for customer shopping. As per Hari Vasudev, Chief Technology Officer of Walmart US, the enhanced Sparky tool would reorder goods, help in scheduling events, and use computer vision technology to recommend recipes according to food items kept within the fridge of the customer.
In addition to this, Walmart is developing an Associate super agent for store associates as well as corporate associates, while a Marty super agent will be developed for suppliers, sellers, and advertisers. Similarly, a Developer super agent would enable the testing, creation, and rollout of AI tools.
