SeededGrasp splits language reasoning from grasp generation and hits 78% real-world success
Yang Xu · hf · 2026-07-23
SeededGrasp is a language-guided grasping framework for complex scenes that separates semantic reasoning from geometric execution.
- A VLM predicts a seed point from the language instruction.
- A lightweight grasp-generation model then uses that seed to produce the final grasp, avoiding expensive end-to-end training.
- The authors release the first multi-embodiment tabletop grasping dataset, with more than 2.5M grasps in cluttered scenes.
- Results show 72% success in simulation and 78% in real-world experiments, outperforming prior baselines.
More from Embodied
- Humanoid robots move from lab demos to real household appliance handling — CurieuxExplorer · 2026-07-23
- Anchor-Align lifts xArm7 real-robot success from 28% to 54% — Dwip Dalal · 2026-07-23
- iFlytek-backed robot startup bets on world models, body awareness and less real-robot data — 量子位 · 2026-07-23
- EgoSteer trains a dual-dexterous-hand system on 9.6K hours of human hand video — 机器之心 · 2026-07-23
- Musk says Tesla is heading into a major autonomy and robotics push — XFreeze · 2026-07-23
- Figure appears ahead of Optimus on autonomy, but demand for humanoid labor still looks huge — markjeffrey · 2026-07-23