Researchers have developed a novel inference-time framework to enable language-driven grasp models to adhere to negative constraints, such as avoiding specific parts of an object. This method, which combines Sequential Monte Carlo with classifier-free guidance, steers sampling away from forbidden regions without requiring any negation-specific training data. The approach utilizes a frozen 3D part-grounding model to identify these forbidden areas from language instructions. To evaluate its effectiveness, a new benchmark called NegGrasp was created, featuring paired positive and negative instructions with constraint-aware metrics. The proposed method significantly reduces the violation rate compared to existing baselines while also improving both constraint-aware and physical success rates. AI
IMPACT This research could improve the precision and safety of robotic manipulation in complex environments.
RANK_REASON This is a research paper detailing a new method for AI grasping models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
