Dexterous AI, Transparent Object Grasping, and the Race to 810K Robot Units

News

Sharpa's MoDE-VLA Framework Achieves 73% Apple Peel Rate with Multimodal AI Hands

Peeling an apple requires continuous force adjustment and real-time object tracking, which is exactly what most robots fail at. Sharpa's modular AI approach, routing vision, touch, and force signals through specialized expert modules, is a meaningful architectural step toward deploying robots in unstructured kitchen and assembly environments.

Research

HEAPGrasp Outperforms GraspNeRF and VGN on Transparent and Reflective Objects at 96% Success Rate

Traditional depth-sensor methods score 52-68% on transparent and glossy objects, which creates real bottlenecks in logistics and manufacturing. HEAPGrasp's silhouette-based approach using a plain RGB camera is a drop-in upgrade that cuts camera travel 52% and execution time 19%, making it practical to retrofit existing lines.

Analysis

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

250% year-over-year growth in 2025 shipments signals the market is past the proof-of-concept stage and entering real deployment cycles. China holds 85% of hardware shipments, but Western firms still lead on software and AI, so the competitive split will shape where value accrues over the next five years.

Tools

AGIBOT's Genie Sim 3.0 Unifies Simulation, Data Generation, and Benchmarking for Embodied AI

The biggest drag on physical AI development is fragmented tooling and expensive real-world data collection. Genie Sim 3.0 generates synchronized multimodal training environments from text or image inputs, which lowers the barrier to building and evaluating robot policies at scale.