VLM-Distilled Next-Best-View Planning
Learning active perception policies from vision-language model preferences.
Overview
This project investigates learning next-best-view (NBV) policies for robotic grasping from offline vision-language model (VLM) preferences. Rather than querying a VLM during deployment, a lightweight NBV policy is trained to imitate VLM viewpoint rankings, enabling efficient active perception.
System Overview
The proposed framework consists of three stages: (1) simulation data collection, (2) VLM-based preference learning, and (3) real-time deployment. A lightweight NBV policy is trained from offline VLM preferences and replaces expensive online VLM reasoning during robotic grasping.
VLM Preference Learning
During offline training, GPT-4o ranks candidate viewpoints according to how well they reveal graspable regions of the target object. These viewpoint preferences are used to train the proposed NBV policy, eliminating the need for VLM queries during deployment.
Experimental Results
The proposed framework was evaluated in simulation using a Franka Emika Panda robot.
- Simulation: 79% grasp success.
- Runtime: 2.6× faster view selection than Information Gain-based NBV.
- Inference: No VLM queries required during inference.
Status
Manuscript in preparation.