Will we have chatGPT moment for Robots?
Snapshot from The Silicon Valley Humanoid Summit.
This week, the Humanoid Summit brought together innovators, investors, and thought leaders at the iconic Computer History Museum in Silicon Valley. Over two days, attendees witnessed groundbreaking advancements in humanoid technology, with keynotes, panels, and live demos showcasing the rapid acceleration of the humanoid robotics ecosystem 🤖
There were dozen of unknowns that experts and humanoid enthusiasts were trying to answer but a very central question that emerged during the event was: Will humanoid robots have their “ChatGPT moment” — a tipping point that transforms how we perceive and integrate robots into everyday life? From discussions on real-world applications like companionship and space exploration to insights on capital allocation and mass production, the summit explored the road ahead for humanoid robotics and I was honored to be part of this event that will be one of the first in the history of Humanoids that attempted to answer that question. The summit undoubtedly underscored how AI-powered robots are poised to redefine industries and human interactions but when it will happen and if at all, it’s still yet to be learned. The energy and curiosity in the room were palpable, setting the stage for transformative innovations that could change the world.
🤖 Building the Robot Foundational Model
Speaker: Sergey Levine, Physical Intelligence
Sergey Levine delivered an inspiring keynote on the evolution of foundational models for robots, drawing parallels between historical milestones like the first phone call in 1876, the first internet message in 1969, and the first robotic foundational model API call in 2023. He introduced the concept of Vision-Language-Action (VLA) models, which go beyond describing images to performing complex tasks based on visual inputs. Levine highlighted the development of π0 (pi zero), a foundational model trained on 10,000 hours of diverse data and fine-tuned with just 20 hours of task-specific data. This efficient fine-tuning process enables robots to perform tasks like box stacking, laundry folding, and even recovering from mistakes. He shared insights on tests with Astribot, which performed coffee-making tasks using only cameras for vision — no touch sensing — demonstrating adaptability to human instructions during execution.
Levine also broke down the anatomy of training foundational models:
Pre-training: Using massive datasets to develop large language models (LLMs) and vision-based models.
Post-training alignment: Leveraging supervised fine-tuning to adapt the model for specific tasks.
With 3.5 billion parameters and a two-week training timeline for vision-language models (VLMs), Levine emphasized the potential of these technologies to revolutionize robot capabilities. However, speed and precision remain bottlenecks, as errors slow task execution and demand near-perfect performance. The transformative potential of foundational models in enabling robots to perform practical, human-like tasks with increasing efficiency and reliability but there is still a lot of work to do to be able to scale it.
For more details, visit pi.website/blog/pi0. You can see several cool demonstration videos like laundry, clothes folding, coffee making tasks and more.
🤖 Humanoids in Space
Panelists: Shaun Azimi, Robotics Technical Discipline Lead, NASA Johnson Space Center, Luis Sentis, Professor, The University of Texas, Allison Thackston, Senior Manager, Robotics, Blue River Technology
Moderator: Gerrit De Vynck, Tech Reporter, The Washington Post
This engaging panel explored the transformative role of humanoid robots in space missions, highlighting their potential to assist with repairs, cooperative assembly, and hazardous tasks, reducing risks for human crew members. Panelists shared the importance of designing humanoids specifically for microgravity environments, noting the challenges and opportunities of integrating these robots into space operations.
Shaun Azimi shared insights on how robots are being tested for manipulation experiments on the ground, with the goal of enabling them to inspect and repair the International Space Station (ISS). These experiments include making robots mobile and teleoperable, extending their functionality to inspect and repair both on Earth and in orbit.
Luis Sentis discussed the development of Valkyrie, one of the first humanoid robot equipped with elastic actuators designed for outdoor operations rather than inside the ISS. New actuators are under development to further enhance its capabilities, enabling tasks in extreme environments beyond Earth’s atmosphere.
Allison Thackston emphasized the critical role of robotic collaboration in building and maintaining infrastructure in space, stressing the importance of intuitive human-robot interaction for teleoperated and autonomous missions.
It’s amazing that the Humanoid robots are also, or mainly (?!), about to revolutionize space exploration, from assisting astronauts to tackling the most dangerous tasks in space, paving the way for deeper human-robot partnerships in the final frontier. One of the first applications should be in the environments that are not safe for people and it seems like the work here is already ongoing!
🤖 Humanoids at Home
Speaker: Bernt Øivind Børnich, Founder and CEO, 1X.
Bernt Bornich is definitely one the most visionary founders of our century. He shared his journey of bringing humanoid robots from concept to commercialization, highlighting the critical shift from prototyping to scalable deployment (vertical integration). While humanoid robots have existed for more that 25 years, Bornich emphasized the importance of timing in their adoption — balancing technical advancements with safety, cost-effectiveness, and societal integration.
Bornich discussed the evolution of 1X’s robots, starting with EVE, launched in 2020, and now transitioning to NEO, their most advanced humanoid. NEO stands out with a lightweight (30 kg/66 lbs), safe, and energy-efficient design powered by tendon-based actuators rather than traditional gears. This innovative approach enhances reliability, reduces energy consumption, and ensures safety during human interaction. With a height of 5’4 ft, 4 hours of operational life, and vertically integrated manufacturing, NEO seems to be scalable and ready for deployment.
1x initially deployed around 80 EVE robots in logistics as pilot programs. However, recognizing the slow pace of enterprise adoption, 1X shifted focus to consumer markets, where it claims that faster deployment, task diversity, and tolerance for failure would make home environments more promising opportunity. Bornich stressed that humanoids need to do more than perform tasks — they should become cultural icons, driving bottom-up societal pressure for widespread adoption. 1X is paving the way for humanoid robots to become a cost-effective and culturally integrated part of our daily lives, bridging the gap between technical innovation and consumer readiness.
🤖 Bimanual Musculoskeletal Humanoids with Artificial Muscles: Redefining Human-Like Dexterity
Speaker: Dhanush Radhakrishnan, Co-Founder and CEO, Clone Robotics
This one was like from real sci-fi movies! Dhanush Radhakrishnan explored how Clone Robotics is redefining human-like dexterity by developing bimanual humanoids equipped with artificial muscles that mimic human skeletal muscles and looks so real!!! You just have to see this live! These artificial muscles are designed to replicate human-level performance, enabling smooth, quiet, and natural movements.
Radhakrishnan highlighted the importance of achieving not just functional capabilities but also natural behavior in robotic operations. By focusing on the integration of artificial muscles, Clone Robotics is pushing the boundaries of humanoid design, bridging the gap between mechanical efficiency and human-like interaction.
This session shed light on the future of dexterous humanoids, paving the way for robots that can seamlessly integrate into human environments with performance and behavior that closely resemble their biological counterparts.
🤖 Humanoid Robots and the Path to Real-World Skills
Speaker: Jonathan Hurst, Co-Founder and Chief Robot Officer, Agility Robotics
Jonathan Hurst, co-founder of one of the pioneers of humanoids - Digit, delved into the challenges and breakthroughs in training humanoid robots for real-world skills, emphasizing the critical role of simulation as the foundation for developing advanced control and planning policies. He explained that real-world environments alone are insufficient for training robots, particularly for complex, whole-body dynamic control tasks where no expert examples exist.
Hurst outlined how expert examples for robot control must be generated in simulation, leveraging specialized hardware and diverse sources like teleoperation, video data, and motion capture. This simulation-generated data can serve as synthetic training material for large models or be directly applied as control policies on physical robots, with further refinement happening through real-world fleet deployment.
He also noted key limitations: while simulations can train certain behaviors, some real-world dynamics — such as balancing, falling, and recovering — cannot be effectively teleoperated and must be approached differently. The shift from single-purpose robots to multipurpose systems is crucial, requiring adaptable workflows and continuous refinement across fleets of robots. The power of simulation in enabling robots to acquire and refine real-world skills, marking a significant step forward in making humanoid robots versatile and effective in dynamic, unstructured environments.
🤖 Humanoid Robotics Foundation Models
Panelists: Sachin Chitta, Director of Robotics Research, Autodesk; Ryan Julian, Research Scientist, Google DeepMind; Remi Cadene, Research Scientist, Hugging Face; Alexander Kernbaum, Founder, Geometric Horizons.
The panel on “Humanoid Robots Foundation Models” brought together experts to discuss the development and deployment of foundational models in robotics. A key point of discussion was the challenge of ensuring reliability in robotic applications. Experts emphasized that, unlike large language models (LLMs) which can afford occasional errors, robotic systems must operate with a high degree of precision to function safely and effectively in real-world environments. This necessitates extensive validation before deployment.
NVIDIA’s Project GR00T, for example, a general-purpose foundation model is designed to enhance humanoid robot learning in both simulated and real-world settings.
Boston Dynamics’ Atlas has integrated advanced AI models to perform complex tasks autonomously, demonstrating significant progress in mobility and manipulation.
These examples illustrate the rapid advancements in applying foundational models to robotics, aiming to equip humanoid robots with versatile and reliable capabilities for various applications and foundation for further development and specialization.
🤖 Standardizing the Future: Developing Frameworks for Safe and Ethical Humanoid Robot Integration
Panelists: Aaron Prather, Director of Robotics and Autonomous Systems Programs, ASTM International, Steve Cousins, Executive Director, Stanford Robotics Center, Stanford University, Kamel Saidi, Program Manager and Group Leader, NIST, Kat Scott, Developer Relations Engineer, Intrinsic
This panel explored the pressing challenges and opportunities in creating frameworks for the safe and ethical integration of humanoid robots. While the panelists agreed that standardization is essential, they acknowledged that it remains a distant goal due to the complexity of developing universally accepted protocols.
Key insights included:
Utility, Usability, and Transparency: These foundational pillars must be addressed to make humanoid robots viable for widespread adoption. Standards must ensure robots are not only functional but also understandable and trustworthy in their operations.
Safety Concerns: Panelists identified stability and tipping prevention as the top priority for standardizing humanoid robot performance. Robots must meet rigorous performance benchmarks to avoid accidents in real-world settings.
Performance Testing: Developing frameworks for consistent performance evaluation is crucial. The robotics community must focus on metrics that ensure reliability and robustness in varied environments.
While the path to standardization is long, the panel highlighted the importance of collaboration between industry leaders, researchers, and regulatory bodies to create frameworks that prioritize safety, utility, and ethical considerations in humanoid robotics. This foundational work will ensure that humanoid robots can be integrated into society responsibly and effectively.
🤖 Dancing with the Robots
Speakers: Dr. Catie Cuan, Robot Choreographer and Postdoctoral Scientist, Stanford University, Ken Goldberg, Distinguished Chair in Engineering, UC Berkeley
Dr. Catie Cuan and Ken Goldberg delivered a captivating presentation on the intersection of robotics and art, exploring how motion can become a medium for creativity and expression. Dr. Cuan, a pioneer in robotic choreography, highlighted how movement design bridges the gap between mechanical precision and human emotion, turning robotic actions into a form of performance art.
Ken Goldberg expanded on how engineering principles, combined with artistic sensibilities, allow robots to interact in ways that resonate with human audiences. The session showcased collaborative projects that explore motion not just as functional behavior but as a language of communication, challenging traditional perceptions of robots as purely utilitarian. A standout discussion focused on how robotic choreography enhances human-robot interaction, making machines more relatable and fostering deeper connections. This session was different and showed how the fusion of engineering and art opens up new possibilities for robotics, offering insights into how robots might seamlessly integrate into human environments, not just functionally but emotionally as well.
I recommend to watch Catie’s inspiring Ted Talk about: Next up for AI? Dancing robots.
So are we closer now to answer if we will have a ChatGPT Moment for Robots?
Maybe an inch, or maybe not even that? 🤔
A Concept Centuries in the Making
The journey of humanoid robots is far from new. From early literary explorations like E.T.A. Hoffmann’s Der Sandman (1816) and Mary Shelley’s Frankenstein (1818) to the first use of the term “robot” in Karel Čapek’s 1920 play R.U.R., the idea of human-like machines has fascinated humanity for centuries. Yet, while concepts like ChatGPT emerged seemingly overnight, humanoid robotics has been an arduous journey — overcoming barriers in motion, dexterity, safety, and societal integration.
Key Insights from the Summit
Data Is Key, but the Bar Is High
Foundational models revolutionized vision and language by scaling data, with breakthroughs like GPT-3 leveraging billions of hours of training. Robotics, however, faces a unique challenge: a scarcity of relevant training data and the complexity of multi-dimensional problems (XD vs. 1D or 2D). Achieving the scale and precision needed for robotics is far more demanding due to the requirement for flawless real-world performance.Simulation as a Catalyst
Simulation seems to be a crucial tool. Robotics experts argued that simulation-based training could generate synthetic data to accelerate progress. Yet, refining these models in real-world fleets will require iterative, long-term efforts — not the overnight scalability seen in language models.Incremental Integration vs. A Tipping Point
Short-Term Impact: Robots are already making strides in specific industries but these are incremental steps, focusing on utility and safety in niche markets rather than mass adoption.
Long-Term Vision: The absence of universal frameworks for safety, performance, and societal integration is a problem. This suggests a gradual integration process rather than an immediate leap.
Cultural Adoption and Bottom-Up Pressure
Robots must move beyond functionality to become cultural icons that create societal demand. Adoption in homes and factories will depend not just on their capabilities but on public perception and trust — a process that could take years.
The consensus across the summit suggests that while humanoid robots may one day reach a “ChatGPT moment,” it is more likely to be a gradual evolution rather than a sudden tipping point. Unlike language models, humanoid robots operate in physical, safety-critical environments, requiring rigorous testing, standardization, and societal acceptance before scaling.
That said, every incremental step — whether it’s improved dexterity from bimanual artificial muscles, better safety frameworks, or scalable manufacturing — brings us closer to a future where humanoid robots become integral to factories, industrial environments, and homes.
The Humanoid Summit 2024 highlighted one certainty: the road to robotic ubiquity may be long, but its foundation is stronger than ever. Whether the “moment” comes suddenly or over decades, the progress being made ensures that humanoids will eventually integrate seamlessly into our world. 🚀
Passionate about HardTech?
Let’s connect: https://bit.ly/3SXPifd
Fantastic and comprehensive write up Paulina.
That was a crazy few days!