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

Researchers have proposed two new frameworks for improving robotic manipulation, addressing issues of inefficiency and generalizability in current systems.

The first framework, GTP-FA (Grasp-Then-Plan with Failure Attribution), is a two-stage approach that generates grasp candidates and performs downstream motion planning conditioned on the selected grasp^[1]. It learns a failure attribution model that generalizes to unseen grasps, producing a stable distribution over failure modes for diagnosis-guided optimization. In experiments, GTP-FA improved task success rates in both simulation and real-robot settings, outperforming base learners in RL, IL, diffusion-policy, and VLA-based settings. Meanwhile, a separate research effort introduced GEAR-VLA, a Vision-Language-Action (VLA) framework designed for generalizable robotic manipulation. GEAR-VLA adopts coarse-to-fine action learning and embodiment canonicalization to address challenges such as unseen objects and background shifts^[2]. The framework achieved state-of-the-art performance on several benchmarks, including LIBERO and RoboTwin 2.0, and obtained a 90.1% success rate on a 6,360-trial universal grasping benchmark with 212 unseen objects.