The Dawn of Smooth Exponential Growth in Robotics
How AI advancements are driving rapid progress in robotics capabilities.
1The Concept of Smooth Exponentials
Dario Amodei's concept of 'smooth exponentials' describes AI progress as an inevitable, steady climb rather than sporadic breakthroughs. Unlike traditional research which follows a 'hero's journey' pattern, AI capabilities advance through aggregate efforts. This phenomenon is now extending to robotics, where multiple parallel advancements create a similar exponential curve. The sheer number of participants ensures progress continues regardless of individual setbacks.
2Ecosystem Maturity in Robotics
The robotics field now mirrors AI's growth trajectory with university labs, startups, and corporate R&D all contributing. China's ecosystem is particularly vibrant, with companies exploring every hardware variation (bipeds, grippers) and application (consumer, enterprise). Open-source vision-language-action models and accessible hardware components lower entry barriers, allowing new players to start further along the capability curve than before.
3LLMs as Robotics Controllers
Frontier models like GPT-6 Astra demonstrate surprising robotics competence without specialized training. Companies like Robocurve are already deploying LLMs for robotic manipulation tasks. This convergence means robotics evaluation will soon adopt LLM benchmarking practices like test-time scaling curves. Publicly verifiable capabilities across domains accelerate progress as improvements in one area benefit all.
4Data Paradigm Shift
Recent breakthroughs show robotics generalization requires different data than previously thought. Instead of millions of robot-specific hours (like Tesla's approach), broad pretraining on domain X (videos, gripper data) transfers to robotics domain Y with minimal specialized data. Open-X Embodiment (2023) pioneered this, followed by UMI grippers (2025) and egocentric video (Nvidia EgoScale 2026).
5Standardized Pretraining Corpus
A reusable robotics training corpus is emerging, hardware-agnostic in nature. Typical 2026 compositions include: millions of egocentric video hours, hundreds of thousands of UMI gripper hours, tens of thousands of teleoperation hours across form factors, and select high-quality hardware demos. This standardization enables faster iteration as new entrants don't need to recreate foundational datasets.
6Capability Projections
Robotics is poised for capability compression - achieving in 12-24 months what took LLMs years. Dyna (2026) and Rhoda (2026) suggest even non-egocentric video may soon contribute to pretraining. The gap between simulation and reality narrows as models demonstrate zero-shot transfer abilities, reducing dependency on physical robot data.
7Chinese Ecosystem Advantage
China's hardware ecosystem provides unique advantages - rapid prototyping, component availability, and manufacturing scale enable faster iteration cycles. Startups can test multiple robot form factors simultaneously. This density of experimentation accelerates the identification of viable approaches while the open-source community shares learnings across teams.
8Implications of Robotic Intelligence
As robotics enters its exponential phase, implications span labor markets, manufacturing, and daily life. The compounding effect of LLM advancements directly benefiting robotics creates a feedback loop. Unlike AI's gradual progress, robotics may experience sudden capability jumps as physical and digital intelligence converge, making this one of the most transformative technological shifts of the decade.
Original article: Eric Jang — The Dawn of Smooth Exponential Growth in Robotics