Vedant Dave

Hey! I'm a Ph.D. student at Cyber-Physical-Systems Lab at Montanuniversität Leoben in Austria and advised by Elmar Rueckert. My research focuses on unsupervised reinforcement learning, robust representation learning under noisy multimodal sensory conditions, and robotic manipulation. I am broadly interested in developing scalable, generalizable representations for embodied agents in real-world, uncertainty-prone environments. Currently, I am a Research Intern in the Machine Learning and Data Science Unit at the Okinawa Institute of Science and Technology (OIST), Japan.

I received my Master's degree in Automation and Robotics from Technische Universität Dortmund in 2021, focusing on Robotics. My thesis, entitled “Model-agnostic Reinforcement Learning Solution for Autonomous Programming of Robotic Motion”, was completed at Mercedes-Benz AG, where I implemented reinforcement learning for motion planning of manipulators in complex environments. Prior to this, I was fortunate to work as a research intern with Leonel Rozo at the Bosch Center for Artificial Intelligence, where I worked on Probabilistic Movement Primitives on Riemannian Manifolds.

 /   /   /   /   / 

Publications and Preprints

Disentangled Diffusion Policy (DisDP) is an imitation learning method that improves robustness to sensor noise and failures by integrating multi-view disentanglement into diffusion-based policies. By decomposing sensory inputs into shared (global) and private (sensor-specific) representations, DisDP preserves task-relevant features while remaining resilient to perturbations. Evaluations on RoboColosseum and Libero show that DisDP matches baseline performance under normal conditions and outperforms them under sensor variations.

EnvoDat is a large-scale, multi-modal dataset designed to enhance robot autonomy beyond urban environments. It includes 26 sequences from 13 scenes, 10 sensing modalities, 1.9 TB of data, and 89K annotations across 82 classes. It is collected in diverse settings like underground tunnels, natural fields, and modern indoor spaces under conditions such as high illumination, fog, rain, and zero visibility. It supports benchmarking for SLAM and supervised learning algorithms and aids in fine-tuning multimodal vision models.

Hilbert Unsupervised Skill Discovery (HUSD) is an unsupervised reinforcement learning method that learns a diverse and distinct set of skills without relying on external rewards. By combining Integral Probability Metrics with \(f\)-divergence, HUSD enhances exploration and ensures skill separability. It outperforms existing algorithms on state-based benchmarks, providing a robust foundation for various downstream tasks.

M2CURL builds on MViTac to improves RL by efficiently integrating visual and tactile representations. It accelerates learning in downstream manipulation tasks.

MViTac is self-supervised approach that integrates vision and tactile modalities using contrastive learning. MViTac utilizes intra- and inter-modality relationships to learn a shared representation space. Experiments show that MViTac outperforms state-of-the-art methods in downstream tasks, which are performed by linear probing.

This paper contains late-breaking results for TacProMPs, which learns and predicts complex arm movements based on tactile responses. Experiments were conducted on the real robot with a wide variety of objects.

TacProMPs learns and predicts complex arm movements based on tactile responses, particularly for manipulating non-uniform objects, demonstrating adaptability diverse grasping scenarios.

This paper introduces a Riemannian formulation of ProMPs for encoding and retrieving quaternion trajectories, enabling full-pose robot motions in operational spaces. This method builds on Riemannian manifold theory and exploits multilinear geodesic regression for estimating the ProMPs parameters.

Industrial Collaborations

Consultant
2022/03-2023/02

Predicting Yield Strength of different materials from production process. Designing Neural Networks and optimization. Found out error in measurement inaccuracy from data analysis.

Academic Collaborations

This approach integrates multi-view disentanglement into diffusion-based policies, enhancing robustness to sensor noise and failures. By decomposing sensory inputs into shared (global) and private (sensor-specific) representations, it preserves task-relevant features while remaining resilient to perturbations.

This research evaluates hole conditioning operations in wellbore drilling, focusing on activities like circulation, reaming, and washing to ensure well integrity. An agent is trained with model-free RL coming from the online data from the real-world scenarios.

Coding and Supervision

A physics-informed neural network combined with the CALPHAD formalism predicts Gibbs energy in alloys by determining the Redlich-Kister parameter using novel descriptors. This method enhances CALPHAD parameterization, expediting materials development and phase stability determination with high accuracy and potential.

Coding and Supervision

This study predicts the floor level from the distribution of TCEs (Li, Be, V, Ga, Ge, Nb, Sb, Te, Ta, Tl, Bi, and REYs) in Vienna's urban aerosol.

Teaching Experience

Reviewing Services

2025: ICML, NeurIPS
2024: AISTATS, ICLR, NeurIPS, ECAI, CoRL, ICRA, IROS, BioRob
2023: ECAI, CoRL, IROS, RAL
2022: ECAI, CoRL, IROS, RAL, Humanoids

Media