Gig Workers, Robotic Skin, and the Unglamorous Infrastructure Powering Humanoid Robots

Research

Sharpa's Apple-Peeling Robot Shows Dexterous AI Manipulation Is Finally Getting Real

A 73% apple-peel completion rate might sound modest, but contact-rich manipulation at this level is a genuine technical milestone. Sharpa's MoDE-VLA framework using modular AI experts for different manipulation phases is the architecture pattern worth tracking here.

HEAPGrasp Achieves 96% Grasp Rate on Transparent and Reflective Objects Using Only RGB Camera

Solving the transparent object problem without depth sensors is a practical unlock for logistics and food automation where glass and shiny packaging are everywhere. The 52% reduction in camera trajectory length means this can retrofit onto existing hardware without a full re-engineering cycle.

Analysis

Humanoid and Quadruped Robot Shipments Hit 53K Units in 2025, Forecast to Reach 810K by 2030

A 250% year-over-year shipment jump in 2025 signals the market has left pilot mode. China accounts for 85% of hardware shipments while Western firms hold the software and AI edge, so the competitive split matters as much as the headline numbers.

Tools

AGIBOT Ships Genie Sim 3.0, a Unified Sim-to-Real Training Stack for Embodied AI

A single platform covering environment generation, multimodal data output, RL pipelines, and standardized benchmarking is exactly the infrastructure gap that has slowed robotics deployment. If the sim-to-real transfer holds up in practice, this becomes a foundational dev tool.

News

How Tokyo University's HEAPGrasp Uses Silhouettes Instead of Depth to See Hard Objects

The shape-from-silhouette approach sidesteps the optical property problem entirely rather than trying to correct for it, which is the right engineering instinct. This kind of architectural rethink is more durable than patch-based fixes to depth sensors.