Benjamin Heltzel

Towards Sim2Real Transfer in Robotic Reinforcement Learning via Proprioceptive Self-Adaptation
Benjamin Heltzel·Leon Sievers·Prof. Berthold Bäuml

We demonstrate real-time proprioceptive adaptation to unknown intrinsic robot dynamics in dexterous in-hand manipulation, paving the way for long-term deployment without re-calibration downtime.

SAC²: Rapid Adaptation to Random Disturbances in Partial Observability via Off-Policy Meta-Reinforcement Learning
Benjamin Heltzel·Leon Sievers·Prof. Berthold Bäuml

An asymmetric LSTM-based meta-RL variant of Soft Actor-Critic enables real-time adaptation by inferring hidden task dynamics from history in POMDPs.

PromptScene: Adaptive Prompt Learning for Open-Vocabulary 3D Instance Segmentation
Benjamin Heltzel, Simon Blessmann, Ayaka Nanri
Mohamed El Amine Boudjoghra·Prof. Angela Dai

Adaptive prompt learning in OpenScene with Mask3D class-agnostic instance grouping enables flexible 3D object recognition without predefined labels.

Dif-fused DINO-tracker: Zero-shot Point Tracking with Video Diffusion Motion Features
Benjamin Heltzel·Dr. Lei Li·Prof. Matthias Niessner

Cross-attention fusion of temporal representations from video diffusion models with DINO features improves zero-shot point tracking coherence.

Seismos Network
Co-founder

We built Seismos to provide global earthquake early warning alerts with ML-based real-time earthquake detection on crowdsourced phone accelerometer data. In the US, the app was also slated to deliver USGS ShakeAlert notifications for CA, OR, and WA.