Physical AI and Humanoids: Breakthrough or Bubble?

The evolution of robotics has reached a pivotal moment with the emergence of cognitive robots, also known as embodied Artificial Intelligence (AI) or physical AI. This fourth generation of commercial robots, following mechanical, agile, and autonomous robots, represents a significant leap in capability and potential impact across various industries.

The cognitive robotics sector, especially humanoid robots, is experiencing unprecedented interest and investment. Major companies like Tesla, with its Optimus project, and startups like Figure, which secured US$675 million in Series B funding, are driving this trend. OpenAI's partnership with Figure to integrate advanced AI models into humanoid robots marks a significant development.

Other key players include Amazon-backed Agility Robotics, planning to grow to produce 10,000 robots annually, and 1X, which raised US$125 million for its NEO android assistant. Chinese companies like Unitree Robotics, UBTECH Robotics, and RobotEra are also advancing humanoid technology.

The automotive industry is embracing this technology, with BMW successfully testing Figure's robot in its plant. Elon Musk's ambitious vision for Tesla's Optimus project further underscores the sector's potential.

These developments, along with projects like Apptronik's Apollo robot and Boston Dynamics' Atlas, highlight the rapid advancement and growing applications of cognitive robotics and humanoid robots across various industries.

This heightened interest is driven by a convergence of technological advancements. New generation sensors and actuators have dramatically improved robots' mechanical agility and perception. Sophisticated computer simulations now allow for accurate digital twinning of robots and their operational environments, accelerating development cycles. These simulations, augmented by Generative Artificial Intelligence (Gen AI), could train the robot to deal with an endless number of virtual tasks and experiences it might encounter in the real world, even if these situations have not yet been practiced in the physical environment. The emergence of Gen AI and Retrieval Augmented Generation (RAG), and Reinforcement Learning with Human Feedback (RLHF) has opened new avenues for personalized and efficient robot training. An example of this is NVIDIA's Project GR00T, an accelerated framework designed to enhance training for general humanoid robots. This framework enables robots to understand natural language and emulate human behavior by observing actions. Furthermore, breakthroughs in accelerated computing have enabled the accommodation of complex workloads, including Artificial Intelligence (AI) training and inference, and multi-sensor orchestration.

The optimism surrounding the technology is also driven by the decreasing costs of these technologies and broader supply chain options, and optimized design and manufacturing techniques enabling the build of durable humanoids at a cost as low as US$30,000 in the near future.

Open-source projects are playing a crucial role in accelerating innovation and attracting developer interest to the cognitive robotics sector. A prime example is Eureka, an open-source model created by NVIDIA that leverages Large Language Models (LLMs) to train robots in complex skills. Eureka's main task is to design reward functions for robot dexterity at superhuman levels, effectively bridging the gap between high-level reasoning (coding) and low-level motor control. This project has demonstrated impressive results, such as teaching robots to perform rapid pen-spinning tricks, showcasing the potential for achieving human-like or even superhuman dexterity in robotic systems.

Eureka utilizes Isaac Gym, a NVIDIA Graphics Processing Unit (GPU)-designed physics simulation platform that can speed up reality by up to 1,000X, significantly accelerating the development and training process for complex robotic tasks. This type of open-source initiative not only democratizes access to advanced robotics development tools, but also fosters a collaborative environment where developers worldwide can contribute to and benefit from cutting-edge research in cognitive robotics. Such projects are instrumental in pushing the boundaries of what's possible in robot dexterity and adaptability, paving the way for more versatile and capable cognitive robots across various applications. Cognitive robots come in various form factors, each tailored to specific applications. While humanoids like Tesla's Optimus and Boston Dynamics' Atlas garner much attention, other forms are equally significant. Quadruped robots excel in challenging terrains, stationary robotic arms dominate in precision manufacturing, and swarm robots show promise in collective problem-solving. Ultra-flexible robots navigate confined spaces, flying robots revolutionize last-mile delivery, and soft robots enable safe human interaction. The diversity in form factors reflects the adaptability and potential of cognitive robotics across different sectors.

The key characteristics of physical AI set it apart from previous generations of robotics. These systems exhibit remarkable versatility, capable of performing a wide range of tasks across different environments. Their ability to learn and adapt quickly, powered by advanced AI algorithms, enables rapid deployment in new scenarios. Personalization through RAG and RLHF allows for tailored responses and actions, enhancing their effectiveness in human-centric environments. In manufacturing and warehousing, their adaptability shines through quick reconfiguration for various tasks. Perhaps most importantly, these robots are designed for advanced human interaction, making them suitable for applications in healthcare, hospitality, and customer service where empathy and communication are crucial.

The potential use cases for cognitive robots are vast and varied. In manufacturing, they promise adaptive assembly lines and sophisticated quality control. Healthcare applications range from surgical assistance to rehabilitation and eldercare. Logistics and warehouse operations stand to be revolutionized by these versatile machines. In agriculture, precision farming and crop management could see significant advancements. Exploration of extreme environments, both in space and deep sea, becomes more feasible. Emergency services could deploy these robots for disaster response and search-and-rescue operations, minimizing human risk in dangerous situations.

However, the path to widespread adoption of cognitive robots is not without challenges. Technology maturity remains a significant hurdle, with issues of energy efficiency, real-time training and reactivity, robust perception, and motor skills still requiring substantial development. The current Bill of Materials (BOM) costs, while reduced, are still too high for mass adoption.

The cost factor is crucial; these systems need to provide better value than human labor at a significantly lower cost to justify large-scale deployment. Regulatory frameworks and ethical considerations, particularly around safety, privacy, and societal impact, need careful navigation. Power consumption and the ability to operate for extended periods in unstructured environments are also key challenges that need addressing. Any interrupted job during working hours could translate into significant lengthened downtime and loss of productivity dollars. Lastly, physical AI and cognitive robotics, especially humanoids, require significant energy for both training and operation. These technologies rely on LLMs and complex simulations, demanding substantial computational resources that could strain the power grid. Large-scale adoption of AI-powered robots would further intensify this challenge. For instance, a manufacturing facility deploying 100 such robots could increase its daily power consumption by an amount equivalent to that of up to 150 average U.S. households. This potential surge in energy demand underscores the need for careful consideration of power infrastructure and efficiency measures as these technologies advance.

Encouraged by the fast development of enabling technologies, market projections for cognitive robotics are bullish, with the most optimistic projections coming from Goldman Sachs indicating that if costs continue to decrease and capabilities improve, the humanoid (a small subsegment of cognitive robotics) could approach US$38 billion by 2035 (10 years from now), with the potential to address 48% of work activities across the economy. Elon Musk was also quoted as saying Tesla has the ability to build and deploy billions of humanoids within the next decade or so.

Although ABI Research foresees a significant growth of humanoids, with early adoption likely in manufacturing, healthcare, and logistics, this market is expected to represent only a small segment of the cognitive and physical robotics’ market. The global market for physical AI is expected to reach a substantial figure by 2030, with the humanoid robot shipments potentially exceeding 200,000 and a revenue figure exceeding US$7 billion.

For industry stakeholders, several strategic imperatives emerge. Prioritizing Research and Development in energy-efficient components and advanced AI algorithms is crucial to overcoming current limitations. Collaborative development efforts, fostering partnerships between robotics firms, AI companies, computing chipset makers, and end users, can accelerate innovation and practical applications. Developing frameworks for measuring autonomy level and establishing comprehensive ethical guidelines for the deployment of cognitive robots is essential to address societal concerns and ensure responsible adoption.

Companies should focus on creating scalable, modular robotic platforms that can be easily adapted across different industries. Emphasis should be placed on human-robot collaboration, rather than full automation, designing systems that complement human workers. Proactive engagement with regulators will be key to establishing standards for testing, deploying, and operating these advanced systems.

While short-term applications are important for market penetration, investment in long-term research for truly transformative technologies should not be neglected. Education and training programs to prepare the workforce for working alongside and maintaining these advanced robotic systems will be crucial for successful integration.

In conclusion, cognitive robotics and physical AI represent a paradigm shift in the field of robotics and AI. Despite the challenges, the potential benefits across various sectors are immense. As technology continues to advance, one can expect to see cognitive robots playing an increasingly important role in shaping the future of work, human-machine interaction, and societal progress. The coming 2 to 3 years will be critical in determining how quickly and effectively this promising technology can be translated from research labs and prototypes into practical, widely-deployed solutions across industries.