**Title: Grasp-Then-Plan with Failure Attribution: A Closed Two-Stage Framework for Precise and Generalizable Robotic Manipulation

Abstract:

The intricate interdependence between grasping and motion planning in robotic manipulation frequently masks the actual root causes of failure, resulting in inefficient trial-and-error cycles. To facilitate efficient long-horizon manipulation tasks, we introduce GTP-FA (Grasp-Then-Plan with Failure Attribution), a task-specific, two-stage framework. This approach first generates grasp candidates and subsequently executes downstream motion planning, conditioned on the chosen grasp. When a manipulation trajectory fails, our system employs a learned failure attribution model capable of generalizing to novel grasps. This model yields a stable distribution across various failure modes, enabling diagnosis-guided optimization. Leveraging these attribution insights, we refine both modules through a diagnosis-driven strategy. Specifically, on the grasping front, we integrate task-level priors and risk penalties into the scoring and optimization of grasp candidates to eliminate unstable or incompatible options. Concurrently, on the planning side, we focus data collection and fine-tuning efforts on high-risk initial states to resolve authentic planning bottlenecks. Our evaluation, conducted across both simulation and real-robot experiments, demonstrates that GTP-FA significantly enhances the performance of base learners in RL, IL, diffusion-policy, and VLA-based contexts, leading to substantially improved overall task success rates.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC