Advocates for developing general-purpose, body-agnostic robotic intelligence, arguing it is ultimately an easier and more scalable approach than creating specialized models tied to specific hardware.
Believes that modern multimodal language models are the most viable path to imbuing robots with common sense, using techniques like 'chain of thought' to unlock their reasoning capabilities for physical tasks.
Posits that the key to overcoming the data scarcity problem in robotics is to deploy useful systems into the real world to create a 'data flywheel' where robots autonomously collect more training data.
Contends that the core research challenge is to fuse the vast world knowledge from generative AI with the capacity of deep reinforcement learning to achieve superhuman performance on physical tasks.
Asserts that progress in generalizable perception systems gives modern robotics a significant advantage over the state of self-driving car development in its early days.
1980s
Levin references this period as the origin of early end-to-end learning for autonomous driving (e.g., the Alvin project) and the technological basis for most industrial robots still in use today.
~2019
Characterizes this as the era before modern VLMs, when robots lacked the common sense needed to make reasonable guesses about physical situations.
~Late 2023
Notes a key inflection point at Physical Intelligence where their models became competent enough to be improved via high-level language supervision, shifting the primary bottleneck from low-level motor control to mid-level scene interpretation.
Present
Describes Physical Intelligence's current model as a VLM with a few billion parameters, a 100ms inference speed, and an architecture based on Google's Gemma with a custom 'action expert' for generating continuous motor commands.
Next 1-2 Years
Expresses a hopeful prediction that a robot capable of performing a useful, real-world task for people will be deployed within this timeframe.
Next 5 Years
Provides a median estimate for when a robot will be capable of acting as a fully autonomous housekeeper.
▶The Thesis of GeneralityApr 2026
Levin's core strategic belief is that developing a single, general-purpose, body-agnostic robotic model is a more efficient and ultimately easier path than creating specialized models for narrow applications. This approach aims to build a foundational intelligence that can be adapted to any physical form, from a multi-fingered hand to a wheeled robot, without changing the core model.
This 'one model to rule them all' strategy is a high-risk, high-reward bet that could create a powerful, defensible platform, but it challenges the conventional wisdom of solving specific, commercially viable problems first.
▶LLMs as the Engine for Embodied Common SenseApr 2026
Levin argues that the breakthrough in large and multimodal language models provides a direct path to solving the long-standing problem of common sense in robotics. By leveraging the vast knowledge within these models and employing techniques like 'chain of thought' reasoning, robots can interpret ambiguous commands and make reasonable guesses about physical situations.
This approach effectively outsources a major part of the AGI problem to the LLM domain, allowing robotics to focus on grounding that intelligence in physical action, significantly accelerating progress.
▶The Real-World Data FlywheelApr 2026
Levin identifies the lack of an internet-scale dataset as a major bottleneck in robotics. His proposed solution is to build systems that are just useful enough to be deployed in the real world, enabling them to autonomously collect vast amounts of data, which in turn improves the model, making the robots more useful and leading to more data collection.
This strategy suggests that the first company to successfully deploy a moderately useful robot at scale could initiate a data feedback loop that creates an insurmountable competitive advantage, similar to what Google achieved in search.
▶Advancing Model Architecture for Physical ActionFeb–Apr 2026
Physical Intelligence's technical approach diverges from methods that treat actions as discrete tokens. Instead, their model, which is a VLM adapted for motor control, uses a diffusion or flow-matching method to generate continuous robot actions, allowing for more fluid and nuanced physical behavior.
This technical differentiation in how the model translates intention to motion could be a key factor in achieving the dexterity required for complex, real-world manipulation tasks that have stumped previous robotic systems.