Benchmarks Bite, RL Gets Faster, and Humanoid Economics Get Tested

Research

RoboBench: A new standard for evaluating LLM-driven robotics

A new benchmark that probably should exist if we're going to build generalizable robot policies. Makes evaluating LLMs for complex, long-horizon robotic tasks systematic, which is a critical infrastructure gap.

Efficient RL for robot manipulation with goal-conditioned priors

This is a direct shot at the sample efficiency problem in RL for robotics. It's moving toward letting a robot figure out a manipulation task in minutes rather than days, which changes the economics of deployment.

One framework for legged robots that do both walking and handling objects

Another piece of the legged manipulation puzzle, moving toward treating locomotion and grasping as a single optimization problem. This kind of unification is essential for truly mobile manipulation.

Domain adaptation techniques for robot sim-to-real transfer

Tackles a core builder problem: how to make a robot that performs well in your messy, specific environment without collecting a mountain of new data. Focuses on the sim-to-real transfer bottleneck.

Generating diverse scenes for robot training data

Another attack on the data problem, this time by generating realistic, varied training scenes automatically. Could help scale up training without needing expensive per-object robot cage time.

Discussion

The real cost and value analysis of humanoid robots

A necessary, potentially sobering look at the actual unit economics of advanced humanoid robots beyond the hype. If you're building a business here, you need to understand these constraints.