Harsh Kumar Agarwal

Five.ai

Cambridge, United Kingdom

harsh19423@iiitd.ac.in

I am an ML researcher working on generative models, multimodal learning, and embodied AI. I am currently a Research Intern at Five.ai in Cambridge, UK. Before that, I completed a Master’s in Artificial Intelligence at ENSIMAG and UFR IM²AG (Grenoble, France), conducting my thesis research as an ML Research Intern at INRIA Grenoble’s RobotLearn team, advised by Dr. Xavier Alameda-Pineda. That work focused on flow-matching models for multimodal, emotion-conditioned motion generation, resulting in EMODY Flow, accepted at the JAV-CG workshop, ACM Multimedia 2026.

Before my Master’s, I worked as a Software/Firmware Developer at IBM (2023–2025), building redundant BMC architectures and firmware tooling for POWER server systems. Prior to that, I conducted undergraduate research at the VISION Lab, IIIT Delhi under Dr. Saket Anand, where I worked on multi-spectral satellite imagery analysis for precision agriculture — resulting in a dataset paper accepted for oral presentation at WACV 2024.

I completed my B.Tech. in Computer Science (BioScience specialization) at IIIT Delhi in 2023 with a GPA of 8.83.

selected publications

ACMMM
EMODY Flow: Emotion-Aware Audio-Driven Full-Body Motion Generation

Harsh Kumar Agarwal, Xavier Alameda-Pineda, and Olivier Perrotin

In Proceedings of the 1st International Workshop on Joint Audio-Video Comprehension and Generation (JAV-CG), co-located with ACM Multimedia, 2026

Accepted

Abs Bib HTML

Embodied conversational agents require synchronized full-body motion (body gestures and facial expressions) that aligns with speech and emotional state. We identify and address a failure of emotion conditioning: a flow-matching model given both a rich audio embedding and a discrete emotion label suppresses the emotion, generating near-identical motion regardless of the specified emotion. We present EMODY Flow, a lightweight ( 35M parameter) flow-matching framework that attaches to a frozen Qwen-3 Omni model and reuses its internal Mimi audio-codecs to condition two parallel DiT generators – one for SMPL-X body pose, one for FLAME facial expressions. A training-time auxiliary emotion classifier restores emotion sensitivity by forcing generated motion to be emotion-identifiable. EMODY Flow sets a new state of the art on BEAT2 gesture quality, with FGD 0.302, Beat Correlation 0.853, and Diversity 24.62 – improving over the best prior results by 26%, 5%, and 62% respectively – and transfers to zero-shot facial animation on TFHP without domain-specific fine-tuning.
@inproceedings{agarwal2026emody, title = {{EMODY} Flow: Emotion-Aware Audio-Driven Full-Body Motion Generation}, author = {Agarwal, Harsh Kumar and Alameda-Pineda, Xavier and Perrotin, Olivier}, booktitle = {Proceedings of the 1st International Workshop on Joint Audio-Video Comprehension and Generation (JAV-CG), co-located with ACM Multimedia}, year = {2026}, archiveprefix = {arXiv}, note = {Accepted} }

WACV

SICKLE: A Multi-Sensor Satellite Imagery Dataset Annotated With Multiple Key Cropping Parameters

Depanshu Sani, Sandeep Mahato, Sourabh Saini, and 5 more authors

In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024

Oral Presentation

Abs Bib HTML

@inproceedings{sani2024sickle,
  title = {{SICKLE}: A Multi-Sensor Satellite Imagery Dataset Annotated With Multiple Key Cropping Parameters},
  author = {Sani, Depanshu and Mahato, Sandeep and Saini, Sourabh and Agarwal, Harsh Kumar and Devshali, Charu Chandra and Anand, Saket and Arora, Gaurav and Jayaraman, Thiagarajan},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages = {5995--6004},
  year = {2024},
  note = {Oral Presentation}
}

ICDMW
Cross-Cultural Recipe Transformation Via Neural Network and Encoder-Based Models

Vibhuti Dhar Khanduri, Mansi Goel, Lakshay Dabas, and 1 more author

In 2025 IEEE International Conference on Data Mining Workshops (ICDMW), 2025

Abs Bib

Developed a neural network classifier to identify cuisine types from recipes. Created an unsupervised ingredient substitution method via semantic embeddings, achieving 90% accuracy in cuisine preservation after cross-cultural transformation.
@inproceedings{khanduri2025cross, title = {Cross-Cultural Recipe Transformation Via Neural Network and Encoder-Based Models}, author = {Khanduri, Vibhuti Dhar and Goel, Mansi and Dabas, Lakshay and Agarwal, Harsh Kumar}, booktitle = {2025 IEEE International Conference on Data Mining Workshops (ICDMW)}, pages = {633--640}, year = {2025}, organization = {IEEE}, }