Isabella Liu
I am a fifth-year Ph.D. candidate at the University of California, San Diego, advised by Prof. Xiaolong Wang. Previously, I was fortunate to be co-advised by Prof. Hao Su. I received my bachelor’s degree in Computer Science and Statistics from the University of Illinois Urbana-Champaign.
My research spans embodied agents, robot learning, and 3D world modeling. I develop autonomous systems that understand and interact with the physical world, learning from experience to improve their capabilities.
Research directions
News
Selected publications
Full list on Google ScholarRecova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation
Recova learns recovery skills in digital twins and refines them through real-world experience to reduce robot failures and human intervention.
ASENA: Self-evolving Agents for Embodied Navigation
ASENA enables coding agents to navigate and solve embodied tasks through online programming and persistent experience, without updating model weights.
Long-WAM: Scaling the Context of World-Action Models
Long-WAM combines autoregressive video pretraining with efficient inference to use longer visual histories for future prediction and real-time robot control.
Long-Horizon Manipulation via Trace-Conditioned VLA Planning
LoHoManip guides short-horizon robot policies with continually updated visual traces for long-horizon planning, execution, and recovery.
In-N-On: Scaling Egocentric Manipulation with in-the-wild and on-task Data
In-N-On combines diverse egocentric human videos with task-aligned demonstrations to improve humanoid manipulation, language following, and few-shot learning.
FreeArt3D: Training-Free Articulated Object Generation using 3D Diffusion
FreeArt3D repurposes pretrained 3D diffusion models to generate textured, articulated objects from a few images without task-specific training.
IMLS-Splatting: Efficient Mesh Reconstruction from Multi-view Images via Point Representation
IMLS-Splatting uses differentiable point-based surfaces to efficiently reconstruct detailed, smooth meshes from multi-view images without extra regularization.
RigAnything: Template-Free Autoregressive Rigging for Diverse 3D Assets
RigAnything generates skeletons and skinning weights for diverse 3D assets in arbitrary poses, without predefined rigging templates.
Dynamic Gaussians Mesh: Consistent Mesh Reconstruction from Dynamic Scenes
DG-Mesh recovers temporally consistent meshes, appearance, and motion from dynamic scenes, including non-rigid deformations and changes in topology.
MeshFormer: High-Quality Mesh Generation with 3D-Guided Reconstruction Model
MeshFormer generates detailed textured meshes in seconds by combining sparse-view images, normal maps, and explicit 3D structure.
TensoIR: Tensorial Inverse Rendering
TensoIR separates geometry, materials, and lighting from multi-view images using compact tensor representations for efficient reconstruction and realistic relighting.
Experience
-
Spring & Summer 2026NVIDIA Research
Research Scientist Intern · Santa Clara
-
Summer 2025Meta Reality Labs
Research Scientist Intern · Redmond
-
Summer 2024Adobe Research
Research Scientist Intern · San Jose
Teaching
- ECE 176 · Introduction to Deep Learning & ApplicationsWinter 2025, Spring 2026
- ECE 285 · Introduction to Visual LearningSpring 2024
- CSE 203B · Convex OptimizationWinter 2022
Teaching Assistant at UC San Diego