Skip to the stories
Robotics & Automation

AI100 Identifies Embodied Intelligence as a Major Robotics Challenge

  • Lucy Smith
  • 382 reads

AI100 highlights physical AI as a major open challenge, with real world complexity still separating robot demonstrations from broad deployment.

AI100 Identifies Embodied Intelligence as a Major Robotics Challenge

Robots have become better at interpreting language, using visual information and learning in simulation. Yet those advances have not removed the central barrier to wider deployment: reliably handling the variety and uncertainty of physical environments. The 2026 AI100 report identifies physical AI and embodied intelligence as one of six open challenges it considers especially important to the field.

The report, titled “AI Goes Mainstream: The Scaling Era and the Rise of Generative AI,” is the third edition of the One Hundred Year Study on Artificial Intelligence. The study examines AI’s development and effects over time, issuing reports every five years. Its international expert committee considers technical, ethical, economic and societal questions, and compares the period since 2021 with the previous panel’s assessments.

From language models to physical action

Robotics has taken up several developments in AI over the past five years. Natural language processing has been integrated into robotic systems, while vision-language-action models bring together visual input, language and actions for training and control. Advances in simulation have also made it possible to train models in virtual environments and then deploy them on physical hardware.

These methods offer ways to connect a robot’s observations with instructions and movement. But a system that performs well in a simulation or a tightly controlled demonstration still has to cope with the physical conditions of its intended workplace. Objects may differ from the examples used in training; surroundings can change; and actions may need to be adjusted as a task unfolds. The report’s focus on embodied intelligence reflects that gap between representing the world in a model and acting successfully within it.

For engineering teams, the distinction matters when judging progress. A convincing demonstration can show that a system has learned a useful behavior under specified conditions. It does not, by itself, establish that the robot can repeat that behavior across varied objects, layouts and operating conditions. Deployment depends on performance across the full task, including the cases that are harder to anticipate in advance.

Deployment remains concentrated

Industrial robots are a major part of the report’s account of physical AI. It states that China deploys more industrial robots than the rest of the world combined and produces about 80% of the world’s humanoid robots. Humanoid robots currently account for only 2% of the market, though the report describes the sector as growing rapidly.

Those figures point to active development, but they do not resolve the broader question of where robots can work reliably. Many established successes remain in constrained settings, such as warehouses, or in narrow applications such as floor cleaning. In these cases, the environment or task can be bounded enough to make the robot’s required behavior more predictable. Work in less structured settings presents a harder test because the system must cope with more variation.

That distinction helps explain why hardware progress and AI capability do not automatically translate into broad use. A robot may have the sensors and actuators needed for a task, and a model may generate plausible actions, while the complete system still struggles with changing conditions. A useful deployment case needs the robot to carry out the task consistently in its actual operating environment, not just demonstrate a capability once.

Three unresolved technical problems

The AI100 panel singles out long-horizon planning under uncertainty, dexterous manipulation of diverse objects and formal performance guarantees as scientific problems that remain open. Each addresses a different part of the reliability challenge.

Long-horizon planning requires a system to choose actions across an extended task while accounting for uncertainty. A plan that makes sense at the start may need revision as the robot encounters new information. Dexterous manipulation raises a related issue at the level of contact: objects vary, and handling them requires more than selecting a broadly suitable movement. The robot must interact with different shapes and properties while adapting its actions to what happens in the moment.

Formal performance guarantees concern how confidently a system’s behavior can be bounded or assessed. Demonstrations can show what happened in a particular trial; guarantees address what can be established about performance more generally. For teams considering deployment, that distinction affects how they evaluate whether a system is dependable enough for its assigned work.

What comes next

Researchers disagree on whether current statistical learning methods can solve these problems or whether robotics will need fundamentally new ideas. That debate is unresolved, and the report presents embodied intelligence as a continuing scientific challenge rather than a problem with an established solution.

The next five years should clarify whether better models and simulation can extend robots into less constrained environments, or whether different approaches are needed to handle planning, manipulation and dependable performance. For now, the practical measure of progress is not simply whether a robot can perform a task under favorable conditions, but how well it copes when the physical world varies.

The full report is available from the One Hundred Year Study on Artificial Intelligence.

Source: Lucy Smith · robohub.org

From the same set