Kavya Shankar

I am a Master's student in Autonomous Systems at Hochschule Bonn-Rhein-Sieg (H-BRS), Germany. I recently completed my thesis on affordance-conditioned robot manipulation from egocentric human video, supervised by Prof. Sebastian Houben and Prof. Hermann Blum.

My interests are in robot learning, 3D perception, and getting modern policies (diffusion, VLAs) to work on real hardware. Before my Master's, I was a research assistant at Fraunhofer IAIS, Vaillant, ubiMaster, and PhenoInspect, working on NLP and computer vision problems.

I am currently building Guido, a mobile assistance robot for hospital patients.

Email  /  CV  /  LinkedIn  /  Twitter/X  /  Github

A mobile robot built to help patients in hospitals — wayfinding, fetching, and a friendly presence. Started in March 2026. ROS 2, with SLAM and Nav2.

A diffusion-based trajectory policy that learns manipulation from egocentric human videos, conditioned on detected functional elements (handles, knobs, buttons). Achieves 63.5% real-world success on drawer/door tasks on the Hello Robot Stretch — a 36 percentage point improvement over the simulation-pretrained baseline.

Multi-view YOLO + SAM pipeline for lifting 2D functional-element detections into 3D point clouds. Voxel accumulation, DBSCAN clustering, plane segmentation, and SAM2 tracking, with iterative mAP optimisation across several lifting strategies.

Integrated the EVA-02 Vision Transformer backbone into an OCT retinal layer segmentation pipeline, replacing the default CNN encoder to improve representation learning on sparse medical-imaging data.

Hybrid time-series model predicting daily question volumes across maths, science and language subjects. Combined classical temporal features with user-activity embeddings to improve over pure ARIMA and pure-ML baselines.

Zero-shot transfer from high-resource to low-resource languages for industrial spare-part recommendation. BERT feature extraction, t-SNE representation analysis, and hybrid retrieval-classification strategies.

Inverse-kinematics controller computing wheel torques from a desired platform force. Direct EtherCAT communication via SOEM, with BLAS/LAPACK for the numerical solve, in C.

Knowledge-distillation model trained on OpenFace visual features across therapy sessions to estimate child engagement during robot-assisted autism therapy. Small-data, high- stakes, per-subject generalisation.