AI Researcher · Honda R&D, Tokyo

Bhuvan
Aggarwal

I build perception systems that ship — from vision-language model research published at ICLR and CVPR to on-device models running in real vehicles and smartphone apps.

ICLR '26 · CVPR '26 · ICCVw '25 · WACV '25  //  IIT Bombay, Institute Silver Medal

Publications Google Scholar LinkedIn GitHub

01 // Peer-reviewed

Publications

ICLR 2026First author

Seeing What's Not There: Negation Understanding Needs More Than Training

B. Aggarwal, A. More, M. Soni, S. D. Bhat — International Conference on Learning Representations

CVPR 2026

AdaPrior: Bayesian-Inspired Adaptive Prior Correction for Long-Tailed Continual Learning

S. D. Bhat, A. More, M. Soni, B. Aggarwal — IEEE/CVF Conference on Computer Vision and Pattern Recognition

ICCVw 2025First author

Model Ensemble to Fuse Geometric and Learning Solutions for Camera Rotation Estimation

B. Aggarwal, A. More, M. Soni, S. Bhat — IEEE/CVF International Conference on Computer Vision Workshops

WACV 2025

PC-GZSL: Prior Correction for Generalized Zero-Shot Learning

S. D. Bhat, A. More, M. Soni, B. Aggarwal — IEEE/CVF Winter Conference on Applications of Computer Vision

02 // Selected work

Experience

AI Researcher

Jun 2023 — Present

Honda Innovation Lab, Honda R&D — Tokyo, Japan

  • Developed the first zero-shot algorithm for negation understanding in VLMs, outperforming SOTA by 20% — accepted at ICLR 2026 and presented at Honda's global HTF 2025
  • Designed a plug-and-play model ensemble for camera pitch/roll estimation — 50% MAE reduction over SOTA across 5 benchmarks, running on-vehicle at 65 FPS — published at ICCVw 2025
  • Engineered a Rider Assistance System for 2-wheelers: a lightweight MobileFormer hitting 0.11° test error at 1/5th the parameters of SOTA, deployed via on-device inference in a smartphone app with real-time collision prediction
  • Built an infrastructure-camera V2X perception pipeline (YOLOv7 + DeepSORT, BEV velocity estimation) feeding real-time collision risk prediction
  • Built a surround-view scene understanding system fusing multi-camera BEV outputs — 3D-transform BEV with slope correction, TensorRT-optimized depth, validated through real-vehicle testing
  • Researching Vision-Language-Action models for end-to-end autonomous driving

Technical Lead — Self-Driving Car (SeDriCa)

Sep 2019 — May 2023

Unmesh Mashruwala Innovation Cell, IIT Bombay

  • Led a 33-member team to Level-4 autonomy at 30 km/h, owning perception, planning, and vehicle integration
  • Designed a multi-task perception network improving inference efficiency by 30%; raised ₹4.5M in technical funding

Computer Vision Intern

May 2022 — Jul 2022

Daikin — Tokyo, Japan

  • Built a document transcription and information-extraction pipeline (10× faster, 93% accuracy), fine-tuning EasyOCR to 91% normalized edit-distance on Japanese documents

03 // Focus

Research areas

Vision-Language Models

Negation understanding, multimodal reasoning, zero-shot generalization, and VLA models for end-to-end driving.

AD Perception

BEV perception, camera calibration, sensor fusion, drivable-area detection in unstructured environments.

Learning Under Imbalance

Long-tailed learning, continual learning, and generalized zero-shot recognition.

Deployment

On-device inference, model compression and quantization, TensorRT / ONNX / CUDA, ROS/ROS2 systems.

04 // Recognition

Honors