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

2026-06-02 · Source: Artificial Intelligence · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

GTP-FA (Grasp-Then-Plan with Failure Attribution) is a novel two-stage framework designed to enhance precise and generalizable robotic manipulation by decoupling grasping and motion planning. This system first generates grasp candidates, then performs motion planning conditioned on the chosen grasp. A core innovation is its failure attribution model, which learns to diagnose failure modes from failed manipulation trajectories and generalizes to unseen grasps, providing a stable distribution for diagnosis-guided optimization. Based on these attribution results, GTP-FA optimizes both its grasping module, by injecting task-level priors and risk penalties, and its planning module, by targeting high-risk initial states through data collection and fine-tuning. Evaluations in both simulation and real-robot experiments demonstrate that GTP-FA significantly improves task success rates across various base learners, including RL, IL, diffusion-policy, and VLA-based settings.

Key takeaway

For Robotics Engineers developing long-horizon manipulation systems, adopting a decoupled grasp-then-plan framework like GTP-FA can significantly improve task success. By explicitly attributing failures to either grasping or planning, you can precisely target optimization efforts, rather than relying on trial-and-error. Consider integrating a failure attribution model to guide your data collection and fine-tuning strategies, enhancing both grasp stability and motion planning robustness in your robotic applications.

Key insights

GTP-FA improves robotic manipulation by attributing failures to optimize distinct grasping and planning stages.

Principles

• Decoupling grasping and planning clarifies failure sources.
• Diagnosis-guided optimization improves robotic task success.
• Task-level priors enhance grasp candidate selection.

Method

GTP-FA generates grasp candidates, then plans motion. A failure attribution model diagnoses issues, guiding optimization of grasp scoring (priors, penalties) and planning (high-risk states).

In practice

• Implement a two-stage grasp-then-plan system.
• Develop a failure attribution model for diagnosis.
• Apply risk penalties to unstable grasp candidates.

Topics

• Robotic Manipulation
• Grasp Planning
• Motion Planning
• Failure Attribution
• Reinforcement Learning
• Diffusion Policy

Best for: Research Scientist, Robotics Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• CreFlow: Corrective Reflow for Sparse-Reward Embodied Video Diffusion RL — CreFlow enhances embodied video generation by localizing RL updates with LTL-based compositional rewards and corrective reflow.
• StepFinder: A Temporal Semantic Framework for Failure Attribution in Multi-Agent Systems — StepFinder efficiently attributes multi-agent system failures by encoding logs into temporal semantic sequences for specialized modeling.
• GroundedPlanBench: Spatially grounded long-horizon task planning for robot manipulation — Jointly planning actions and spatial grounding improves robot task success and action accuracy for VLMs.
• SECOND-Grasp: Semantic Contact-guided Dexterous Grasping — Integrating semantic reasoning with physical stability enhances robotic dexterous grasping performance and adaptability.
• Task Capability Improvement Algorithm for Collaborative Manipulators — Utilizing undesired moments from off-center force application significantly enhances collaborative manipulator task capability.
• From folding boxes to fixing vacuums, GEN-1 robotics model hits 99% reliability — GEN-1 achieves high reliability and improvisation in physical tasks by scaling training data and enabling real-world adaptation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.