Sergey Levine

Sergey Levine is a professor of Electrical Engineering and Computer Sciences at UC Berkeley, and the co-founder of robotics startup Physical Intelligence. His research more or less defined how modern robots learn from data. If a humanoid in 2026 walks via reinforcement learning, or grasps an object via imitation learning, it's standing on Levine's papers.

Why his work matters

Before about 2015, two big approaches dominated robotic learning:

• Classical control — write the math by hand for every robot behaviour. Reliable but inflexible. A robot only does what's been explicitly programmed.
• Reinforcement learning — let the robot try and learn. Theoretically elegant but, in practice, took millions of attempts and barely worked on real robots.

Levine and his Berkeley group (BAIR — Berkeley AI Research) pushed a third approach: deep reinforcement learning combined with imitation learning. Show the robot demonstrations from humans (or simulated experts). Let it improve from there with RL. The combination dramatically reduced the amount of trial-and-error needed.

His 2015 paper "End-to-end training of deep visuomotor policies" was one of the first showing a robot could learn manipulation tasks (stacking blocks, fitting plugs) by mapping camera pixels directly to motor commands — no hand-engineered features in between. That paper is the foundation of what's now called end-to-end robot learning.

What he's working on now (2026)

Levine co-founded Physical Intelligence (sometimes branded as π) in 2024, alongside Karol Hausman and others. The startup's mission: build a general-purpose AI brain that can transfer between any robot body. Backed by Jeff Bezos and others; the company has demonstrated VLA models running on a dozen different robotic platforms.

If you think of the humanoid race as a contest between "build the body" companies (Tesla, Figure, Boston Dynamics) and "build the brain" companies, Physical Intelligence is the leading "build the brain" play.

What to read

• End-to-end training of deep visuomotor policies (2015) — the foundational paper
• RT-2: Vision-Language-Action Models (2023, Google DeepMind with Levine as advisor) — the VLA paradigm
• Physical Intelligence π0 model (2024) — the first cross-embodiment robot foundation model

His students and collaborators include Chelsea Finn (Stanford), Pulkit Agrawal (MIT), Karol Hausman (Physical Intelligence) — together this generation of researchers shapes most of the academic work robotics startups are now commercialising.

Why this matters for the field

For the past decade, the bottleneck in robotics has shifted from hardware to software. The robots can move; the question is whether they can decide what to do. Levine's research arc — and his current commercial bet — is essentially the answer to that question. If general-purpose robot brains are possible, his pipeline of work is the most direct path there.

If you want to actually try the ideas Levine pioneered, reinforcement learning will be Forge 03 in our learning track.