HEAPGrasp, Apple-Peeling AI, and the Simulation Stack Closing the Dexterity Gap
Research
HEAPGrasp Hits 96% Grasp Rate on Transparent and Reflective Objects Using Only RGB Camera
Transparent and shiny objects have been a persistent blind spot for robot graspers, but this silhouette-based approach sidesteps depth sensor limitations entirely. A 52% reduction in camera trajectory and 19% faster execution means it is practical enough to retrofit onto existing systems.
Sharpa's Robot Peels Apples Autonomously Using Modular AI Dexterity Framework
Apple peeling is a useful benchmark because it requires continuous force adjustment and real-time contact sensing across an irregular surface. Sharpa's MoDE-VLA framework, which routes different subtasks to specialized AI modules, doubled success rates on contact-rich tasks compared to baselines.
Tools
AGIBOT's Genie Sim 3.0 Unifies Environment Creation, Training, and Benchmarking for Embodied AI
Fragmented tooling has been a real bottleneck in robotics development, and consolidating sim, data generation, and evaluation into one pipeline cuts engineering overhead significantly. The text-to-3D environment feature is particularly useful for scaling training scenarios without expensive real-world data collection.
Analysis
Global Humanoid and Quadruped Robot Shipments on Track for 810,000 Units by 2030
Shipments hit 53,000 units in 2025, a 250% year-over-year jump, and Chinese manufacturers account for 85% of that volume. The split between China dominating hardware and Western firms leading on software and AI is shaping up as the defining market structure for the decade.
News
Getting a robotic hand to survive 150,000 grasping cycles while staying cool enough for continuous industrial operation solves two problems that have quietly killed previous dexterous systems. The open-sourced TacRefineNet framework and 61 hours of tactile data are the most actionable part of this release for anyone building manipulation systems.
Stay Ahead
Delivered each morning.