6D Object Pose Tracking in Internet Videos for Robotic Manipulation

TL;DR

This work tackles learning robotic manipulation from Internet instructional videos by estimating temporally coherent 6D object poses without exact 3D meshes. It combines semantic CAD-model retrieval with patch-based, category-level pose alignment, monocular depth grounding via an LLM-derived size prior, and robust object tracking to produce smooth 6D trajectories that are retargeted to a 7-axis robot through trajectory optimization. Across YCB-V, HOPE-Video, and in-the-wild datasets, the RGB-only pipeline outperforms state-of-the-art methods and demonstrates real-world robotic imitation in simulation and on a real panda robot, including egocentric EPIC-KITCHENS scenarios. The approach offers a scalable path toward large-scale robotic learning from Internet videos by leveraging retrieval, tracking, and SE(3) trajectory optimization to bridge perception and manipulation.

Abstract

We seek to extract a temporally consistent 6D pose trajectory of a manipulated object from an Internet instructional video. This is a challenging set-up for current 6D pose estimation methods due to uncontrolled capturing conditions, subtle but dynamic object motions, and the fact that the exact mesh of the manipulated object is not known. To address these challenges, we present the following contributions. First, we develop a new method that estimates the 6D pose of any object in the input image without prior knowledge of the object itself. The method proceeds by (i) retrieving a CAD model similar to the depicted object from a large-scale model database, (ii) 6D aligning the retrieved CAD model with the input image, and (iii) grounding the absolute scale of the object with respect to the scene. Second, we extract smooth 6D object trajectories from Internet videos by carefully tracking the detected objects across video frames. The extracted object trajectories are then retargeted via trajectory optimization into the configuration space of a robotic manipulator. Third, we thoroughly evaluate and ablate our 6D pose estimation method on YCB-V and HOPE-Video datasets as well as a new dataset of instructional videos manually annotated with approximate 6D object trajectories. We demonstrate significant improvements over existing state-of-the-art RGB 6D pose estimation methods. Finally, we show that the 6D object motion estimated from Internet videos can be transferred to a 7-axis robotic manipulator both in a virtual simulator as well as in a real world set-up. We also successfully apply our method to egocentric videos taken from the EPIC-KITCHENS dataset, demonstrating potential for Embodied AI applications.

Figures (11)

• Figure 1: Robotic manipulation guided by an instructional video. (a) Given an instructional video from the Internet, our approach: (b) retrieves a visually similar mesh for the manipulated object from a large mesh database, (c) estimates the approximate 6D pose trajectory of the object across video frames, and (d) transfers the object trajectory onto a 7-axis robotic manipulator.
• Figure 2: Overview of our 6D object pose estimation without a known 3D mesh (Sec. \ref{['sec:object_pose']}). Given an input RGB image, our method: (a) detects and segments objects present in the image, (b) retrieves similar meshes from a large-scale object database via patch-based retrieval, (c) estimates the absolute scale of depicted objects in the scene via LLM-based re-scaling, and (d) estimates the camera-to-object rotation $R$ and translation $t$ via alignment of the retrieved (approximate) mesh.
• Figure 3: Qualitative results of our method on the YCB-V dataset. On the YCB-V dataset our method is able to detect objects that are not part of the dataset (such as keyboard in image F and chair in image C), which highlights the benefit of our method of not being restricted to the ground truth meshes.
• Figure 4: Qualitative results of our method on the HOPE-Video dataset. On the HOPE-Video dataset our method reconstructs complex scenes with multiple objects while respecting partial occlusions.
• Figure 5: Qualitative results of our method on the EPIC-Kitchens dataset. Our method is able to detect and align objects from an egocentric point of view. See Figures \ref{['fig:appendix_qualitative_results_epic_kitchens_1']} and \ref{['fig:appendix_qualitative_results_epic_kitchens_2']} for additional examples.