← Research  /  RL² Series

Welcome to the Era of Physical Evals

In software, evals became the bottleneck — and the sharpest teams in AI are racing to build them. In the physical world, that layer doesn't exist yet. That's the whole game.

Jerry Huang · Robotics Center of Silicon Valley · July 2026

Reinforcement learning is saturating every benchmark. In software, that already flipped the bottleneck: the hard part is no longer the model — it's building the evals and environments to train and judge it.

That's the thesis behind Mercor's "evals are the new PRD" and Fireworks' "Production RL." Two of the sharpest teams in AI are racing to build eval infrastructure for software agents. They're right.

Here's what almost nobody is saying: in the physical world, that layer doesn't exist yet. And it's a much harder problem.

Models commoditize. The loop does not. Value in Physical AI accrues to whoever closes the real-world learning loop fastest and cheapest.

Why physical is different

  • You can't clone a workcell. A software agent's environment spins up in a container. A robot's environment is a real cell — real actuators, glare, wear, unit-to-unit variance.
  • You can't auto-grade a folded shirt. Code compiles or it doesn't. "Did the robot place the box correctly?" has no unit test.
  • The trace isn't a log. In software, the trace is ground truth. In the physical world, the trace is a robot that overheated at 4pm and dropped the part.

Robot models will saturate any benchmark too. The barrier to putting robots to work across the economy isn't a bigger model — it's real-world evals, which barely exist for physical tasks.

RL²: Reinforcement Learning in Real Life

It's why we built the Robotics Center around a single loop: deploy → capture failures → evaluate against real acceptance criteria → collect targeted data → redeploy. Real hardware, real failures, continuous learning. Two things follow.

Acceptance criteria are the new PRD — for robots. Whoever encodes what "working" means in a real deployment owns the gate every model version has to pass.

The loop, not the model, compounds. Hand-debugging each deployment failure is a variable cost that never ends. An eval harness turns it into a fixed asset that pays on every version.

Where we actually are

Honestly: no one has software-grade, fully-automated physical eval today. It's a hybrid of automated regression gates and operator-in-the-loop judgment, and full automation is a multi-year build. That's why it's worth building — and why it can't be done from a desk. It needs real hardware, real operators, and a real-world loop.

Software-only companies can't reach this layer. The model labs don't want the physical dirty work. That leaves a wide-open seat: the physical-world version of what Mercor and Fireworks are building for software. That's the seat we're taking.

In software, the trace is ground truth. In the physical world, the workcell is.

If you're training policies that work in the lab but not the workcell — that's the wall we work on.

Read the technical whitepaper → See SVRC Eval Compare notes with us