Macro Micro News Global Pulse. Local Truth.

Why Your Robot Fails at Pouring: The Two-Stage AI Fix for Task-Specific Grasping

05 November 2026 · 2 min read

We compile, generate and translate using Artificial Intelligence from the below given source. Macro Micro News is responsible for its editorial publication.

Article image by David Levêque
Image by David Levêque

Switzerland, Source:

Robotic manipulation has long struggled with a specific disconnect. Traditional systems can easily identify multiple ways to hold an object based on geometry alone. Yet they often fail to select the pose that allows for successful downstream actions. Imagine a robot holding a kettle perfectly but unable to pour because it missed the handle's functional orientation. A new two-stage framework addresses this by decoupling grasp generation from semantic reasoning. This approach significantly improves task-conditioned selection accuracy.

The architecture operates in two distinct phases to solve this problem. In Stage 1, a Vision-Language Model identifies task-relevant regions within an RGB-D observation. This guides segmentation and the generation of plausible 6-DoF grasp candidates via GraspNet. The result is a set of proposals that are both geometrically sound and tied to specific functional areas. Stage 2 then employs a structured VLM critic to rank these grounded candidates. It evaluates them based on compatibility with the intended task rather than relying solely on geometric scores. This bounded approach prevents the model from generating unconstrained poses. Instead, it selects from a validated set of candidate IDs.

Extensive evaluation across 60 full-system physical trials demonstrated an overall task success rate of 63.3%. The tests covered tasks like random pick-and-place, hammer hitting, and pouring. Controlled replay experiments revealed significant variations in performance depending on the VLM choice. Gemini 3 Flash Preview emerged as the top performer for Stage 1 localization. It achieved 95.0% accuracy with the lowest latency. For Stage 2 ranking, GPT-5.6 Sol demonstrated superior reliability. It secured a 92.9% Top-1 task-valid rate and showed the strongest consistency in repeated inferences.

This study highlights that physical task completion and model-specific behavior capture different aspects of system robustness. Sensitivity analysis showed that changes in upstream candidate evidence could alter final grasp selections. This emphasizes the need to evaluate both stages jointly. By separating semantic isolation from candidate auditing, this framework provides a clearer basis for assessing VLM-based robotic systems. It moves beyond simple success metrics to understand where and why decisions succeed or fail. This methodology offers a scalable path for integrating high-level reasoning into low-level robotic control. It ensures that robots not only grasp objects but do so in a way that fulfills complex human instructions.