Jiuyi (Joey) Xu徐久乙

Education & Experience

Education

Research Experience

Industry Experience

Publications

Projects

CureWM

CureWM

World ModelsRobot LearningFailure Data

World Models Dream of Success: robot world models used for policy evaluation, planning and synthetic data generation must tell successful actions apart from failures, yet across four released checkpoints from two architecture families they respond weakly to action changes and predict success-like outcomes for verified failures. CureWM repairs released checkpoints without changing their architecture or training objective: it perturbs successful demonstrations across a severity grid, verifies the outcomes by execution in simulation or on hardware, and fine-tunes on the resulting failures and surviving successes alongside nominal data. On 484 held-out LIBERO failure counterfactuals, optimism falls from 80% after fine-tuning on the official data to 38% on average across four CureWM models; on a physical robot arm, success-like failure predictions drop from 90% to 33%. With failure counts, successful replay data and training budget matched, counterfactual failures yield a success–failure value gap of 0.124, versus 0.014 for freshly collected on-policy failures.

PreDE

PreDE

WAMQuantizationOffline Prediction

PreDE (Predict Before You Deploy) is a policy-calibrated framework that predicts whether a quantization configuration will degrade a world action model's task performance before it is ever run in closed loop. From the closed-loop outcomes of a small development set, it calibrates two thresholds on offline action deviation, then accepts, rejects, or defers each new configuration using only a fixed observation log and cached reference actions. Across five WAMs (Cosmos Policy, UVA, mimic-video, RynnVLA-002 and Fast-WAM) and four benchmark settings, task losses cannot be explained by bit width alone or by a shared deviation threshold. On 28 held-out configurations from two policies, PreDE issued 21 decisions (75% coverage) before observing any closed-loop outcome, all matching the observed labels. In 450 Franka Research 3 trials, every configuration assigned to a high-deviation group showed significant degradation, while W4A4 on the real robot delivered a 1.37× action-query speedup and about 44% lower peak memory.

VLAQuantBench

VLAQuantBench

VLAQuantizationPTQ Benchmark

A controlled closed-loop evaluation of post-training quantization (PTQ) for VLA models: 409 runs and 94,574 simulation episodes covering OpenVLA-OFT, π₀, π_0.5 and X-VLA on LIBERO, with X-VLA further evaluated on SIMPLER, CALVIN and VLABench. Holding a shared round-to-nearest quantizer fixed while varying numerical format and layer scope, it shows that quantization sensitivity is neither monotone nor additive — under uncalibrated W4A4, expanding a π_0.5 action-head subset from 126 to 167 quantized layers raises success from 7.0% to 70.5%. For OpenVLA-OFT, protecting a single 28,672-parameter output projection restores near-baseline success. The results identify concrete, recipe-dependent precision assignments rather than universal layer-sensitivity rules, and are complemented by real-kernel memory and latency measurements and a physical Franka Research 3 study.

FastVLA

FastVLA

VLAToken CompressionTraining-Free

A training-free, task-aware, and temporally adaptive visual-token compression framework for efficient closed-loop VLA inference. At each step it scores token importance via task-to-vision attention and sets the compression rate from the importance-distribution uniformity, measured by an inverse-Simpson metric with online sliding-window calibration. Across OpenVLA, OpenVLA-OFT and CogACT on LIBERO and SIMPLER it reduces LLM-backbone FLOPs by 35–59% and action latency by 9–32% (1.09×–1.46× speedup) while maintaining task success; on a physical Franka Research 3 robot it cuts action-prediction latency by 20% (1.25× speedup) at comparable success.

LowDiff

LowDiff

DiffusionEfficient SamplingImage Generation

A cascaded diffusion framework that generates increasingly higher-resolution outputs, but replaces the usual stack of stage-specific models with a single unified model that progressively refines images from low to target resolution. It reaches comparable or better quality with far fewer high-resolution sampling steps and applies to both pixel-space and latent-space diffusion. Across CIFAR-10, FFHQ and ImageNet (conditional and unconditional) it delivers 50–80% throughput gains — e.g. FID 1.94 on conditional CIFAR-10 and FID 1.59 on ImageNet-256 (LightningDiT-XL/1) — while maintaining quality.

Honors & Service

Awards

• 2026 NVIDIA Academic Grant2026
• 2025 ASCE Outstanding Reviewer2026
• 2025 ASCE Best Data Award2025

Academic Service — Reviewer

News

• Aug 21, 2026Completed my Robotics internship at Prime Robotics.
• Jun 8, 2026Received an NVIDIA Academic Grant.
• May 18, 2026Started as a Robotics Intern at Prime Robotics.
• May 11, 2026Passed my Ph.D. qualifying exam.
• Apr 10, 2026Recognized as an Outstanding Reviewer of 2025 by the Journal of Computing in Civil Engineering.
• Mar 18–21, 2026Presented "EgoGen: Generating Egocentric Images on Construction Sites Using Diffusion Model" at CRC 2026.
• Dec 7–10, 2025Presented "GenVision: Enhancing Construction Safety Monitoring with Synthetic Image Generation" at WSC 2025.
• Aug 2025Received the 2025 ASCE Best Data Award.
• Apr 24, 2025Gave an invited talk "Multiscale Conditional Diffusion on Image Generation" at Colorado School of Mines, hosted by the IEEE Student Branch.
• Aug 2024Began the Ph.D. program in Robotics at Colorado School of Mines, advised by Dr. Yangming Shi.
• Jul 28–31, 2024Presented "An Open-Vocabulary Framework for Efficient 2D–3D Visual Data Annotation in Indoor and Outdoor Built Environments" at i3CE 2024.
• May 2024Earned an M.S. in Computer Science from the University of Southern California.
• Feb 11–13, 2024Presented "A Hybrid 2D–3D Zero-shot Framework for Photogrammetric Data Segmentation" at Geo Week 2024.
• Nov 27–Dec 1, 2023Presented "USC-ICT Semantic Terrain Points Labeling System" at I/ITSEC 2023.
• Jun 2023Started as a student research intern at USC-ICT (Geospatial Terrain Research Group), supervised by Dr. Meida Chen.
• Aug 2022Began the M.S. program in Computer Science at the University of Southern California.

Contact