My Avatar

李炫毅 / Xuanyi Li

VLA · World Models · Robotic Systems

VLA · World Models · Robotic Systems

Building the intelligence layer for physical robots.

I lead work on vision-language-action systems, world-state modeling, and production autonomy stacks that connect perception, planning, prediction, and action.

2026 - Present

Building VLA and world model systems for embodied robots.

01 · XLab · Embodied Intelligence

At XLab, my work focuses on embodied intelligence: using VLA policies, world-state modeling, multimodal perception, and closed-loop action generation to build robot systems that can understand, predict, and act in physical environments.

XLab embodied intelligence media reference.

2024 - 2026

VLA / XPlanner for robotic decision and action.

02 · Robot Policy

At XPeng Motors, I work on vehicle-side VLA/XPlanner systems: route-video-to-trajectory modeling, large-model scaling, dynamic interaction, and complex-scenario action generation. I think of this as robot policy learning under real product constraints.

2021 - 2024

Model the world before taking action.

03 · World Models

My prior work at DJI Automotive focused on BEV perception, dynamic object detection, tracking fusion, occupancy-style scene understanding, and 4D annotation loops. These are the ingredients for world-state modeling in deployed robot systems.

DJI world model and perception visual
DJI dynamic perception and world model visual
DJI embodied world model media visual
DJI embodied intelligence media visual
DJI perception and planning media visual

2018 - 2021

3D vision as the sensorimotor substrate.

04 · Perception Foundation

I built and maintained practical stereo and depth systems, including X-StereoLab with 600+ stars and 100+ forks. Stereo matching, active stereo, RGB-D understanding, and road-structure perception form the lower-level grounding for robot intelligence.

3D vision and sensorimotor substrate visual

Mission

Build general, reliable physical AI robot systems.

My goal is to build robot intelligence that can generalize across physical environments, deploy at real-world scale, and continuously improve through closed-loop data, world models, and self-evolving iteration.

Physical AI robot system mission visual

Visitors

Global visits

-- Total views
-- Visitors