NVIDIA · Agora

Selected work

Representative Papers

LidarDM: Generative LiDAR Simulation in a Generated World

Apr 03, 2024IEEE International Conference on Robotics and Automation

Existing methods for generating realistic, layout-aware, physically plausible, and temporally coherent 4D LiDAR video in autonomous driving simulation remain inadequate. Method: This paper introduces the first generative 4D LiDAR world model tailored for driving scenes. It employs an integrated framework that leverages latent diffusion models for 3D scene generation, jointly models dynamic agent motion, and captures spatiotemporal (4D) point cloud sequences; LiDAR video is then synthesized via differentiable sensor rendering. Contribution/Results: Our approach pioneers driving-semantic-guided 4D LiDAR generation, uniquely ensuring layout consistency, physical interpretability, and temporal coherence. Quantitative and qualitative evaluations demonstrate significant improvements over prior art in realism, temporal continuity, and structural fidelity. The generated LiDAR sequences effectively support downstream perception model training and evaluation.

39 citations5 influentialRead paper

On Ray Reordering Techniques for Faster GPU Ray Tracing

May 04, 2020ACM Symposium on Interactive 3D Graphics and Games

Secondary rays in GPU-accelerated ray tracing exhibit poor spatial locality and low SIMT execution efficiency, limiting overall performance. Method: This paper proposes a hardware-agnostic ray reordering framework centered on an Endpoint-Estimated Key—designed specifically for secondary rays—to enhance spatial coherence with minimal computational overhead. The approach integrates wavefront path tracing scheduling, cooperative utilization of RTX hardware traversal kernels, and GPU memory access locality modeling, without relying on shader-specific implementations. Contribution/Results: Evaluated on modern GPUs, the method achieves 1.3–2.0× end-to-end ray tracing speedup. Reordering significantly accelerates software-intensive stages (e.g., intersection computation and shading preparation); although hardware traversal incurs inherent overhead, the net performance gain remains unambiguous. This work establishes a novel, portable paradigm for general-purpose GPU ray reordering—enabling efficient, implementation-independent acceleration across diverse ray tracing pipelines.

18 citations3 influentialRead paper

Intuitive Surgical SurgToolLoc Challenge Results: 2022-2023

May 11, 2023

To address the challenge of real-time, robust surgical instrument localization in minimally invasive robotic-assisted surgery (RAS) video streams, this work introduces SurgToolLoc—the first large-scale, multi-view, multi-scenario benchmark dataset with pixel-level mask annotations. We further propose a novel evaluation protocol emphasizing both cross-center generalizability and real-time inference (≥30 FPS). Methodologically, we integrate instance segmentation and keypoint detection with temporal modeling (ConvLSTM/Transformer), domain adaptation, and weakly supervised learning. Our best-performing model achieves 92.4% mAP@0.5 on the test set while maintaining an inference speed of 36 FPS—substantially outperforming conventional template matching and early CNN-based approaches. The solution has undergone rigorous preclinical validation across multiple surgical scenarios. By providing a reproducible, scalable, end-to-end framework for visual instrument localization in RAS, this work establishes a new standard for benchmarking and advancing vision-based surgical navigation systems.

17 citationsRead paper

GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization

Jan 08, 2026arXiv.org

This work addresses a critical limitation in multi-reward reinforcement learning, where the existing method GRPO suffers from collapsed advantage estimates due to uniform normalization across combined reward components, thereby degrading training signal resolution and leading to convergence difficulties or premature failure. To overcome this, we propose Gradient Descent Policy Optimization with Decoupled Normalization (GDPO), which normalizes each reward component independently to preserve their relative differences, thus enhancing the accuracy and stability of advantage estimation. Theoretical analysis reveals, for the first time, the normalization flaw of GRPO in multi-reward settings. Extensive experiments demonstrate that GDPO consistently outperforms GRPO across diverse tasks—including tool use, mathematical reasoning, and code generation—achieving significant improvements in accuracy, error rate, and adherence to format constraints, while exhibiting superior performance and generalization capability.

14 citations1 influentialRead paper

Learning to Discover at Test Time

Jan 22, 2026

This work proposes TTT-Discover, a novel approach that introduces test-time training (TTT) to scientific discovery by leveraging online reinforcement learning to optimize large language models during inference. Unlike conventional AI methods that rely solely on the generalization of pretrained models and struggle to autonomously identify optimal solutions at test time, TTT-Discover focuses on generating a single high-quality solution rather than improving average performance. Built upon the open-source model OpenAI gpt-oss-120b and augmented with a customized search subroutine and the Tinker API, the method achieves highly efficient and low-cost optimization. It sets new state-of-the-art results across diverse domains—including mathematical theorem proving, GPU kernel design, algorithmic competitions, and single-cell denoising—with all findings validated by domain experts or competition organizers at a cost of only a few hundred dollars per task.

11 citations1 influentialRead paper

Recent publications