Embodied artificial intelligence (AI) — AI that interacts with the physical world through robots, autonomous vehicles, or other hardware — is making significant strides beyond research laboratories. However, as highlighted at the recent ATxSummit tech conference in Singapore, the path to widespread adoption remains steep. Industry leaders stressed that while robots are becoming more capable thanks to advances in hardware, simulation, and sensors, broader deployment will hinge on solving critical challenges related to reliability, safety, cost, data scarcity, and stronger governance standards.
Embodied AI differs fundamentally from software-based AI, which processes digital data and generates outputs confined to screens or servers. By contrast, embodied systems must perceive, reason, and act in the unpredictable, noisy, and often hazardous physical world. This distinction adds layers of complexity — robots require not only intelligent algorithms but also robust mechanical design, real-time sensor fusion, and fail-safe mechanisms to avoid causing harm. The field brings together robotics, computer vision, natural language processing, and control theory, and it has captured the imagination of researchers, investors, and governments globally.
William Dally, chief scientist and senior vice-president of research at Nvidia, highlighted a key breakthrough: using AI to enable robots to perform tasks they have not been explicitly programmed for. He demonstrated a humanoid robot assembling a model car from a simple text prompt, showcasing how robotic foundation models can translate visual inputs into motor actions autonomously. Such models, trained on vast datasets of robot interactions, allow machines to generalize across tasks and environments, moving beyond the narrow, pre-scripted behaviors of traditional industrial robots.
Yet, the industry is still in its infancy, cautioned Yutaka Matsuo, a professor at the University of Tokyo. “We are not in the full adoption phase at the moment,” he said. Matsuo emphasized that better architectures, algorithms, data, compute resources, cost efficiency, and safety systems are all necessary before robots can operate reliably outside controlled settings. He pointed out that today’s robots often stumble in unfamiliar environments, struggle with object manipulation in clutter, and require significant human oversight.
The timeline for embodied AI deployment was laid out by Om Nalamasu, senior vice-president and chief technology officer at Applied Materials. He described 2024 as a proof-of-concept phase, 2025 as a year of demos, and 2026 as a period for pilots. Nalamasu noted that the industry has shifted from asking whether such systems can be built to how they can be deployed safely and reliably at scale. He stressed that for robotics to scale, systems must achieve lower latency, greater energy efficiency, and cost-effectiveness. Sensors will be critical — robots depend on sensor fusion, combining data from cameras, LiDAR, touch, and inertial measurement units — to understand and respond to the physical world accurately.
Data remains another major constraint. Unlike software-based AI models trained on internet-scale text, real-world robotics data is scarce, expensive to collect, and often requires human teleoperation or careful simulation. Nalamasu added, “We need to be thinking about standards, interoperability and the governance model,” pointing to the need for common frameworks that allow different robots and systems to work together and be certified for safety.
Real-world deployments
Despite the challenges, commercial deployments are emerging in structured environments. Zhao Yuli, chief strategy officer at Galbot, reported that the company has deployed more than 1,000 robots in China across humanoid-operated stores, logistics facilities, and pharmaceutical chains. Galbot uses a mix of real-world and synthetic data to train its systems. However, generalization remains a significant hurdle. Zhao noted that while robots perform well in known or semi-structured environments, they struggle in unfamiliar settings. As a result, Galbot is focusing first on semi-structured scenarios where robots can learn from real-world deployments incrementally.
Suthen Thomas Paradatheth, chief technology officer at Grab, highlighted that robotics — unlike software — involves significant physical concerns and “does not have near-zero marginal cost.” He explained that each robot requires hardware, maintenance, supply chains, sensors, and physical integration with buildings. Updating a robot fleet is far harder than updating cloud software, especially if new sensors or hardware modifications are needed. Grab conducts extensive testing before scaling: “Before scaling to hundreds of robots, we make sure we crack it first in simulation and with a few robots, while building a data flywheel to monitor, learn from and improve each deployment,” said Paradatheth. Establishing trust is also essential; Grab’s autonomous vehicle pilot clocked 40,000 kilometers and involved months of testing, stakeholder engagement, and community consultation before public deployment. Paradatheth stressed that the approach should begin with the customer problem, not the technology: “Fall in love with the customer problem, but don’t fall in love with the solution set.”
Public-private collaboration
The panellists also highlighted the need for public-private collaboration. In China, Zhao noted that government support has helped create testbeds, strategic projects, and long-term funding for embodied AI. In Japan, Matsuo pointed to the AI Robot Association (AIRoA), an open data initiative targeting 100,000 hours of robotics data for researchers and companies developing robotic foundation models. Such datasets are vital for training generalizable skills, and sharing data across organizations can accelerate progress while reducing duplication of effort.
Safety standards will be critical because embodied AI can affect the physical world directly — unlike purely digital AI systems, robots can cause physical harm if they fail. Matsuo suggested that Japan and Singapore, with their strong regulatory environments and advanced robotics ecosystems, could help shape global standards for safety, interoperability, and governance. Nalamasu added that progress in robotics will not be linear: advances in hardware, software, and data will reinforce one another in a “multiplicative” way. For example, better simulation tools enable faster algorithm development, which in turn informs hardware design, leading to more capable and affordable robots.
Dally stressed that embodied AI will only become practical if intelligence can run efficiently on the device itself. “We need to run them on real robots, and these can’t be tethered with an umbilical cord back to the datacentre. They have to be carrying the intelligence on them,” he said. This will require more efficient chips, software frameworks, and model architectures that can handle complex perception and control tasks within tight power and latency budgets. Nvidia, for instance, is developing specialized processors and edge computing platforms aimed at bringing large language models and vision transformers onto robots.
The promise of embodied AI remains significant. Speakers pointed to ageing populations, labour shortages, healthcare, manufacturing productivity, and city operations as key areas where the technology could deliver value. In the near term, industrial and semi-structured environments — factories, warehouses, hospitals, retail stores — are likely to lead adoption. Over time, autonomous robots are expected to move deeper into public spaces and homes, performing tasks such as delivery, cleaning, caregiving, and personal assistance. However, that future depends on sustained investment in research, cross-sector partnerships, and the development of robust safety and governance frameworks that earn public trust.
Source: ComputerWeekly.com News