Robotics Roundup: VC Money Piles Into Humanoids While Nobody Solves the Data Problem

Analysis

Scaling Laws for Robot Manipulation Data Are Different From LLMs

This paper argues robot policies don't benefit from scale the same way language models do, because the data distribution shift between sim and real (or lab and deployment) dominates everything else. If true, this is the single most important thing for founders to internalize before burning cash on data collection ops.

Research

Whole-Body Control for Humanoids Using Learned Dynamics Models

Instead of hand-tuned controllers, this approach learns the dynamics model directly and plans through it, which could mean faster iteration cycles when hardware changes. Practical if you're building humanoid firmware and tired of re-tuning gains every time a new actuator ships.