SILICON VALLEY — Ken Goldberg, a leading robotics scholar at the University of California, Berkeley, pointed to OpenAI's latest artificial intelligence model, GPT-6 Astra, as he argued that a breakthrough shift could also come to physical AI, the field that combines AI with the physical world. Just as the launch of ChatGPT in November 2022 opened the era of generative AI, expectations are building that Astra could dramatically raise the performance of humanoids, or human-like robots.
Speaking at K-AI Tech Week in Santa Clara, California, on the 11th, Goldberg said, "I hope for a ChatGPT moment in robotics as well." He added, "When will this happen in robotics? Very interesting things are happening in the research field."
Goldberg cited Astra, which ChatGPT developer OpenAI unveiled on the 3rd. "Astra appears very strong in the area of agentic robotics," he said. "The landscape is changing quickly, and it will spread like wildfire." A prominent scholar in the field, Goldberg is also a co-founder of Ambi Robotics, a two-armed robot company whose name comes from the word "ambidextrous."
Goldberg's remarks point to the arrival of an era of agentic robots, in which AI agents combine with hardware so that robots reason and plan on their own and carry out actual tasks across multiple steps. Astra is the first AI model to receive a "Critical" security rating and is regarded as having state-of-the-art capabilities in areas including cybersecurity and science. Nvidia Chief Executive Jensen Huang praised the model on social media on the 6th, writing that it took four years to get from ChatGPT to Astra, that AGI had arrived, and offering congratulations to the OpenAI team.
Jay Chui, founder of Robocurve, which measures robotics benchmarks, said on the 5th that Astra scored 95% on robot control tasks. Compared with Anthropic's Fable 5.1 model, which scored 40%, Astra delivered far better results while using 6.2 times fewer output tokens and costing 2.3 times less. On social media, Goldberg wrote that he was glad to see Astra post rapid benchmark gains in agentic robotics, and that research on new harness design could improve performance further. A harness refers to the execution environment that sets the goals, procedures, tools and verification standards an AI needs to carry out tasks reliably.
A scholar who has spent 40 years on robots is taking an interest in AI models because he expects AI to make up for the limits of data-driven learning. Goldberg believes robotics research divides into model-free and model-based approaches, and that satisfactory results require combining the two appropriately. Model-based work is the traditional engineering approach of building mathematical models from mathematics, physics and science to control robots. Model-free work, by contrast, relies on learning from data. His argument is that while the industry places blind faith in model-free methods, model-based methods must be pursued alongside them.
Goldberg pointed to a laundry-folding robot he had developed as an example. It was trained with a model-free approach and applied no mathematical model. "My wife looked at the robot and said she didn't like it," he said. "Model-free tends to plateau at around an 80% success rate and does not reach the 95% to 99% success rates that industry demands. That is the dilemma."
Large-scale training can produce one-dimensional large language models and two-dimensional vision models, Goldberg explained, but robots occupy a domain of far greater dimensional complexity. "If it is a robot hand with 22 degrees of freedom, that is 22 dimensions to control, and adding the other hand makes 44 dimensions," he said. "Add the arms and legs, and you have to control more than 50 degrees of freedom, and every angle across all of them has to be perfectly synchronized for the robot to work reliably." Degrees of freedom, or DOF, in a robot hand refers to the number of joints that can move independently.
Data for developing humanoids is itself scarce. "We have vast amounts of data to train vision systems and language systems, but we have no data for robotics. It isn't on the internet either," Goldberg said. His point was that while the words and sentences used to train LLMs can be found on the internet and in written material, the human physical experience needed to train robots is hard to obtain. For that reason, the robotics industry has been fitting factory workers in emerging markets such as India with smart glasses to gather data.
Goldberg compared Google's Waymo and Tesla, the two pillars of self-driving cars. He said Tesla has run vast amounts of driving tests but lags Waymo's performance when accidents are taken into account, and that the reason is Waymo's appropriate use of both data-driven learning and AI models. In other words, it pursues model-free and model-based research together. "Waymo uses a lot of data, but it also uses model-based control theory," he said. "Everyone is excited about model-free theory, but sometimes model-based methods get the job done."
Coding technology has played a large role in the evolution of AI models. "All of you have seen the remarkable effects of coding over the past years and months, with tools like OpenClaw, Claude and NemoClaw," Goldberg said. "This is an agentic process." He said he got chills when he realized that coding could be the way to connect the two mindsets — model-based and model-free — into one, and that this led him to develop a technique called Graph as Policy, or GaP, with Nvidia. GaP is a technology in which multiple agents control each part of a robot through a graph, rather than relying on a single AI model.
