Fei-Fei Li

StereoPolicy: Enhancing Robotic Manipulation with Implicit Geometric Reasoning via Stereo Vision

StereoPolicy is a novel framework designed to improve robotic manipulation by directly utilizing synchronized stereo image pairs. It employs a Stereo Transformer with cross-attention to implicitly capture geometric and disparity cues, enhancing reasoning without explicit 3D models. This approach has demonstrated consistent performance gains over various baselines in both simulated and real-world robotic manipulation tasks, underscoring the efficacy of stereo vision in bridging 2D visual understanding with 3D geometric reasoning for robotics.

Shifted Twisted Quantum Affine Algebras: Structure, Representations, and Classification

This paper introduces shifted twisted quantum affine algebras, a twisted version of shifted quantum affine algebras, derived from Drinfeld's current presentation. These algebras admit a triangular decomposition and their structure is independent of the specific coweight pair, depending only on the total shift. The research establishes a rationality theorem for Cartan currents, enabling the classification of simple objects via rational $\ell$-weights. Furthermore, it develops a fusion product compatible with $q$-characters and characterizes finite-dimensional simple modules, providing a $q$-characters formula for representations by relating them to twisted quantum affine Borel algebras.

Closing the Planning Gap: Self-Exploration and View Graph Distillation for VLM 3D Navigation

This paper introduces "view planning," the ability of Vision-Language Models (VLMs) to predict view changes from camera movements and plan multi-turn actions in 3D space. Current VLMs exhibit a planning gap, failing to compose view-action knowledge across multiple steps, particularly as distances increase. The proposed iterative framework enhances VLMs by alternating self-exploration with view graph distillation, forming a comprehensive view graph from all trajectories and converting it into supervised tasks to mitigate sparse rewards and improve planning capabilities.

ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop