Generative World Models
Action-conditioned, multi-view video models for controllable, reproducible, long-horizon simulation.
Researcher · Engineer
Chaoda Zheng
I build generative models that imagine how the physical world may unfold, then use those imagined futures to help intelligent agents reason and act.
01 · About
From understanding geometry
to simulating futures.
My research began with 3D perception: point-cloud tracking, segmentation, occupancy, and geometry-aware representations. Today, I focus on action-conditioned world models, causal video generation, VLA systems, and closed-loop simulation.
I was the first author and a core contributor to X-World. Alongside hands-on research, I define problems, design experiments, and mentor early-career researchers across both product and academic projects.
2,000+
Scholar Citations2 Oral
CVPR / ICCV Papers1 Spotlight
NeurIPS Paper02 · Research Interests
What I work on
VLA & Embodied Decision
Multimodal action decoding, latent future reasoning, and reinforcement-learning post-training.
3D Perception
Object-centric occupancy, point-cloud tracking, segmentation, and geometry-aware representation.
Scalable Model Systems
Efficient autoregressive inference, sequence parallelism, and high-throughput closed-loop evaluation.
03 · Featured Work
X-World
Arxiv 2026 · First Author
Controllable ego-centric multi-camera world models.
A seven-camera action-conditioned world model for stable 30+ second generation and VLA closed-loop evaluation. I led causalization, few-step distillation, long-horizon post-training, and the first simulation inference pipeline.
- I 30+ sec stable controlled generation
- II 50 → 4 denoising steps per block
- III ~49× throughput gain over bidirectional baseline
Arxiv 2025 · Corresponding Author
FutureX
An end-to-end driving policy that reasons through possible futures in latent space before choosing an action.
ICCV 2023 Oral · Second Author
LATR
Transformer-based 3D lane detection from monocular images, bridging image features and lane geometry in 3D space.
04 · Selected Publications
Papers
-
2026
X-World: Controllable Ego-Centric Multi-Camera World Models Arxiv · First Author↗
-
2025
FutureX: End-to-End Driving via Latent Chain-of-Thought World Model Arxiv · Corresponding Author↗
-
2024
Object-Centric Occupancy Completion Augments 3D Object Detection NeurIPS · First Author↗
-
2023
A Motion-Centric Paradigm for 3D Single Object Tracking TPAMI / CVPR Oral · First Author↗
-
2023
LATR: 3D Lane Detection from Monocular Images with Transformer ICCV Oral · Second Author↗
-
2022
2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds ECCV · Co-First Author↗
-
2021
Box-aware Feature Enhancement for Single Object Tracking on Point Clouds ICCV · First Author↗
05 · Academic Community
Service
ECCV 2026
Area Chair
Top Venues
Reviewer
CVPR · ICCV · ECCV · NeurIPS