Jiang Chaokang
World Models · Generative Driving · 3D/4D Perception
A 7-camera surround world model, compressed from 35 denoising steps to 4 for up to ~13.9× faster rollouts, now driving a sensor-level closed-loop simulator at Bosch — alongside perception that runs on production autonomous-driving platforms.
Born in 1998 in Henan, China. I received my Master of Engineering in 2023 through a joint program (2020–2023) between the China University of Mining and Technology and Shanghai Jiao Tong University, where I was fortunate to be advised by Prof. Hesheng Wang and Prof. Yanzi Miao, and to build multi-year collaborations with Dr. Guangming Wang and Dr. Jiuming Liu, both now at Cambridge. At the SJTU IRMV Lab I worked on a 4D auto-labeling system and a lawn-mowing robot — my first experience of shipping perception onto real hardware rather than a benchmark. I then spent 2023.06–2025.03 at PhiGent Robotics in Beijing before joining Bosch China.
Explore
Scene flow, LiDAR odometry, 3D/4D perception, world models, and generative driving.
A filterable list where every card stays two sentences: one for the problem, one for the core contribution. Each entry carries verified PDF, arXiv, DOI, code, and project-page links.
Autonomous-driving, robotics, auto-labeling, deployment, and data-generation systems.
Problem, what I built, and the measured result — then an interactive logic map of the architecture. Enterprise work is sanitized and kept high-level.
Notes on models, systems, deployment, and research practice.
Durable notes rather than news: model design, system trade-offs, deployment lessons, and the judgement calls behind them.
Selected Publications
Rolling a driving world model forward on dense BEV rasters is expensive, which is what keeps world models out of streaming use.
Contribution. It runs the diffusion flow over a compact vector graph of agents and lanes instead of a dense raster, so rollout cost scales with the number of scene elements rather than with resolution — cheap enough to generate the future as a stream.
Large-scale LiDAR registration is dominated by the cost of attention over hundreds of thousands of unordered points.
Contribution. It keeps the registration in 3D but moves the attention onto a range projection, so a 2D transformer carries the heavy computation while the point geometry stays exact — no voxel downsampling, no hand-built descriptors.
Scene-flow networks return one motion vector per point and no indication of when that vector should not be trusted.
Contribution. It refines flow as a hierarchical, coarse-to-fine diffusion process and reads per-point uncertainty directly out of that process — the journal extension of the CVPR 2024 method.
Selected Projects
A generative simulation platform built on Cosmos-Transfer2.5: a 7-camera surround world model, real-map scenario generation, and a 4-step distilled sampler that makes surround rollouts fast enough to sit inside a closed loop.
A two-level driving simulator: a vector world model that decides what the traffic does, and a sensor-level pipeline that decides what the cameras see — reconstruction, generated traffic, and a mask-guided video editor in one closed loop.
A controllable surround-view driving generator: 3D boxes and maps become spatial conditions, text / reference frames / lanes / calibration become condition tokens, and a UNet diffusion backbone turns them into cross-camera-consistent 4V / 7V / 11V video.
Type to search across publications, projects, and blog notes.