This can achieve impressive results. Boston Dynamic’s agile humanoid robots use this approach,⁹ for instance — but typically only for highly constrained tasks. Advances in AI mean these modules are increasingly being replaced by more flexible, learned models. Deep learning already powers Simultaneous Localization and Mapping (SLAM) in autonomous vehicles and drones,¹⁰ and helps sorting robots pick out and manipulate objects.¹¹ Progress is being made in applying AI to locomotion and planning too.¹²

Transformers – the algorithm behind large language models – promise to greatly expand AI’s use in robotics.¹³ LLMs themselves provide a more natural way to interface with robots via language.¹⁴ But transformers also power new multimodal models that can work with multiple data types.¹⁵¹⁶ At the cutting-edge, this is driving a shift from modular architectures to “end-to-end” models that simultaneously solve perception, control, and planning.¹⁷¹⁸ The latest generation of vision-language-action models are able to tackle a wide variety of tasks on a range of robotic hardware platforms.^{19 20} There is also growing interest in the use of diffusion models, which power AI image generators, for robot motion planning.²¹

Recent advances in Physical AI have accelerated the development of general-purpose robotic capabilities by connecting foundation models to perception, planning, and control. However, Physical AI should not be regarded as a complete scientific explanation of embodied intelligence. Many of the fundamental challenges identified by Cognitive Developmental Robotics remain unresolved, including developmental learning, embodied understanding, intrinsic motivation, social cognition, shared meaning, and lifelong adaptation. Physical AI should instead be used extensively as a powerful engineering and experimental tool for investigating these questions.

The ability to draw correlations between different data sources, including language, vision and motion, could allow robots to develop a deeper understanding of the physical world and their place in it — something referred to as embodied intelligence.^{22 2324} This could overcome AI’s “symbol grounding problem”,²⁵ the ability to link the learned word or “symbol” of an object, such as “cup”, to the object’s real-world context. It could also allow AI to develop a sense of “self”^{26 27} although a comprehensive theoretical understanding of consciousness remains a challenge.

Moreover, current models are data- and power-hungry and incapable of learning on the fly. Their huge size means they must be stored on the cloud and communication delays between servers and the robotic hardware can be too long for fine-grained motor control. Their decision-making is also inscrutable, raising safety concerns. Breakthroughs in lifelong learning and interpretability will be crucial in the long run.^{28 29}

The severe shortage of real-world interaction data has also renewed interest in combining learned models with established robotics and engineering methods. Rather than replacing classical robotics, foundation models are increasingly likely to complement physics-based models, inverse kinematics, feedback control, SLAM, modular architectures, and other forms of “Good Old-Fashioned Engineering” needed for speed, reliability, interpretability, and safety.

A promising direction is therefore not a simple transition from modular systems to end-to-end models, but the construction of hybrid architectures that combine learned representations and policies with engineered skills and explicit control structures.

Novel manipulator designs are allowing robots to carry out more delicate tasks than were previously possible.³⁰ But they are still a long way from replicating the dexterity of the human hand. But they are still a long way from replicating the dexterity of the human hand. In particular, tactile sensing is still unable to provide the precise feedback crucial for handling soft and delicate objects or operating safely and socially alongside humans.^{31 32}

However, advances are being made and the latest robotics research platforms now include high-density tactile skins, improved fingertip sensors, and biologically-inspired cameras.³³ The field of soft robotics could also provide solutions, in the form of active materials that can sense and actuate simultaneously.³⁴ In particular, electronic skins that can provide multimodal sensory feedback are advancing rapidly.^{35 36} Artificial muscle technology is a promising option,^{37 38 39} and the ability to rapidly transition between soft and rigid states could make it safer to deploy around humans than conventional hardware.

There are more prosaic concerns, though. Both the computer hardware required to run AI and the mechatronics that enable robots to move are power-hungry, severely limiting run times for mobile robots. Better battery technology and more efficient chips will be crucial before robots can be more widely deployed. Robotic hardware is also expensive, though prices are falling.⁴⁰ Cheap open-source robotics platforms in particular promise to democratise access to advanced hardware.⁴¹ The rapid emergence of commercially available humanoid platforms equipped with high-performance actuators, including quasi-direct-drive systems, is lowering barriers to entry and substantially improving dynamic locomotion. Shared platforms reduce the need for each research group to design and manufacture complete proprietary systems, improve reproducibility, and enable researchers to focus more heavily on perception, learning, coordination, strategy, and Human-Robot Interaction. This transition was particularly visible at RoboCup 2026. Leading teams increasingly used commercially available Chinese humanoid platforms, and the first full 11-versus-11 humanoid robot soccer match was played by two German teams using robots designed by Booster Robotics.^{42 43} The event represents both a major technical milestone and a shift from competitions centred primarily on constructing robot hardware toward competitions increasingly centred on the intelligence, coordination, and autonomy implemented on shared platforms.

More broadly, there is often a lack of imagination when it comes robotic body plans. Most research is focussed on robotic arms, quadrupeds or humanoid robots, but the technology doesn’t need to be constrained to form factors already found in nature. Evolutionary robotics, which adapts robot designs to specific challenges,⁴⁴ and robots that can dynamically alter their configuration hold considerable promise.⁴⁵

The rapid diffusion of humanoid and other mobile robots also raises questions that cannot be addressed solely by improving the efficiency of individual components. If robots become pervasive in homes, workplaces, and public environments, their aggregate requirements for energy, materials, manufacturing, maintenance, and computation will become a significant societal issue. Biologically inspired morphology, distributed control, local body-environment interaction, and the strategic use of ambient energy may therefore become important principles for sustainable robotics.

Transformer-based AI presents both opportunities and fresh problems. Transfer learning — the ability to take a model trained on one kind of robot and deploy it on another — has been a long-standing challenge as even small changes in environment or robot configuration can throw models off.⁴⁶ Transformers can train on data from multiple robots to create more general policies that work across varied embodiments and environments.⁴⁷ However, training one of these models requires colossal amounts of data.

One potential workaround involves training models in simulations before porting them over to real-world robots.⁴⁸ However, many of the tasks planned for robots, such as handling soft and delicate objects, remain hard to simulate.⁴⁹ Designing virtual worlds is complicated and expensive, requiring sustained effort from multi-disciplinary teams.⁵⁰ But high-fidelity simulators are becoming increasingly accessible,^{51 52} and there have been significant advances in techniques for transferring skills learned in virtual environments to real-world robots.⁵³ Emerging “world models” that can generate physically-realistic 3D environments from scratch could also become a powerful tool for training robots.⁵⁴

Pooling data collection efforts will be crucial going forward, and there are promising efforts to create massive open robotics datasets.^{55 56}There will also be a growing focus on more data-efficient training approaches.^{57 58} Data shortages are likely to ease as more robots are deployed, though this will benefit the private companies that build them more than academic researchers.

Despite the rapid development of foundation models, real-world robotic data remain fundamentally scarce compared with internet-scale text and image data. Collecting physical interaction data is slow, costly, hardware-dependent, and subject to safety and reliability constraints. Simulation, world models, teleoperation, shared datasets, and data collected during productive robot operation will all contribute, but none provides a complete solution on its own.

This continuing bottleneck strengthens the case for hybrid approaches. Data-driven learning should be combined with physical models, engineered skills, structured control, and prior knowledge so that robots can achieve useful and reliable behaviour without requiring unrealistic quantities of real-world experience.

At the most basic level, there has been progress on “human-aware” navigation algorithms that allow robots to safely occupy the same space as people.⁶⁰ Collaborative robots, or “cobots”, are also capable of simple interactions like object handovers.⁶¹ But more sophisticated forms of human interaction are facilitated by deep reservoirs of implicit knowledge about social etiquette, and efforts to imbue this understanding and corresponding behaviour in machines remain rudimentary.⁶²

LLMs present a promising opportunity to chip away at these challenges by allowing humans to interface with robots through natural language.⁶³ There is even tentative evidence that they have limited ability to model the mental states of humans.⁶⁴ The outputs of these models can be inconsistent and unreliable though, which makes them unsafe to deploy in many real-world situations.

Understanding how humans interact with robots, both physically and psychologically, and what their expectations of the technology are, is also crucial.^{65 66 67} Better understanding of how integrating robots into groups of humans impacts social dynamics is also needed.⁶⁸

Human-Robot Interaction is increasingly recognised as culturally situated. Robot behaviours designed around individual autonomy, independence, efficiency, and task execution may not transfer directly to societies that place greater emphasis on cooperation, empathy, relationality, solidarity, and collective adaptation. Cultural values shape not only how people interpret robots, but also which robot roles, behaviours, and forms of autonomy are considered desirable or legitimate.⁶⁹ A simple contrast between “Western” and “Eastern” approaches should nevertheless be avoided. Recent European discussions surrounding solidarity, co-shaping, and the societal implementation of AI regulation indicate a growing interest in relational and mutually adaptive approaches. Conversely, East Asian societies contain diverse and changing attitudes toward autonomy and technology. Comparative HRI should therefore investigate how these perspectives can complement and correct one another rather than treating them as fixed cultural opposites.

There is already evidence that the use of robots in social care can have unintended consequences that harm those they are supposed to help.⁷⁰ And given the prevalence of bias in AI training data, robots could replicate existing patterns of discrimination.^{71 72} Developing socially adept robots which feel empathy, exhibit moral reasoning and act accordingly may also require them to be able to feel pain, raising many ethical questions.⁷³ Establishing legal guidelines and frameworks to address the rights and responsibilities of autonomous robots will be required.

This suggests a gradual shift from treating robots primarily as functional instruments or substitutes for human labour toward considering them as potential collaborators, companions, and co-inhabitants of shared social environments. Such a transition requires attention not only to technical capability, but also to dependency, power, responsibility, cultural diversity, and the institutions that structure long-term human-robot coexistence.

Future HRI may increasingly be understood as a process of co-shaping. In this view, robots do not merely execute predefined human goals, and humans do not simply adapt to fixed machines. Instead, human practices, robot capabilities, social expectations, and governance arrangements develop together through sustained interaction.

Embodied intelligence should consequently be understood at multiple levels. Physical embodiment concerns the relation between control, morphology, sensing, and the immediate environment. Social embodiment concerns interaction, mutual adaptation, shared practices, and cultural meaning. Ecological embodiment concerns the aggregate relationship between robotic systems, infrastructures, energy use, material resources, and planetary sustainability.