Nils Kuhn

Research interests

My research interests broadly span robot learning, with a focus on vision language action models and diffusion-based approaches for tasks involving mobile manipulation, long horizon planning, dexterous manipulation or super human performance. Specifically, I am curious about hierarchical approaches and approaches that support continual learning.

• 

  Mobile manipulation

• 

  Dexterous manipulation

• 

  Continual Learning

Publications

• 

  Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them

  Why do Vision-Language-Action models struggle with contact-rich manipulation tasks that demand precise physical interaction? While prior work has focused on fixing these failures with force-augmented architectures and training-time regularizers, we dig into the root causes. We identify two distinct failure modes: precision failures, stemming from a mismatch in flow-matching policy training, and force failures, arising from the unique structure of force signals. By addressing each with a targeted mechanism, we introduce FACT—achieving a 66% average success rate across five contact-rich tasks, compared to 41% for the best prior baseline, validated across almost 2,500 real-world rollouts.

• 

  Validation-gated continual learning for anomaly detection in satellite telemetry

  Satellite health monitoring relies on anomaly detection, but traditional deep learning models are static, struggling when data drifts or is scarce, like at the start of a new mission. Our framework periodically adapts anomaly detection models using continual and active learning, while ensuring every update is validated before deployment. Benchmarked on the European Space Agency's satellite telemetry dataset, our approach delivers a 92.9% event-wise F1 score. A 14.3% improvement over static baselines, while using less than 5% of the original training data.

Projects

• 

  Severance: When Unlearning Crime Unlearns Survival

  We reproduced and extended Lee et al.'s UNDO framework for robustifying LLM unlearning against relearning attacks by testing it on Pythia 160M with the TOFU benchmark. We test unlearning within a multi agent simulation with Mistral 7B. In our multi agent crime simulation, our adapted method drove criminal behavior to 0% across all 80 rounds. However, agents then failed to survive, choosing the correct life saving action only 20% of the time versus 54% for the baseline model. These results suggest that current unlearning techniques are suppressing target behavior while eroding adjacent reasoning that the model needs to function.

• 

  Continuous-Time Diffusion Policies for Visuomotor Control:
  A Stochastic Calculus Perspective

  We reframed diffusion policies as controlled SDEs (Variance Preserving / Variance Exploding / Critically Damped Langevin Dynamics) for visuomotor control. Then, we implemented multiple samplers and compared robustness on the PushT benchmark. Our experiments found that VP-SDE performs best and most consistently (≈0.78-0.80 success).

• 

  GeoVision: Fine-Grained Urban Geolocation in San Francisco via Distribution-Aware Visual Models

  GeoVision predicts where a street photo was taken within San Francisco using a Vision Transformer (StreetCLIP) plus custom geolocation heads. A 31x31 grid classifier reaches 66.8% top-1 accuracy, while a probabilistic Gaussian head achieves ~600m mean localization error and provides uncertainty estimates. The project includes map-based visualizations and attention rollouts to show what cues the model uses when localizing images.

• 

  MARIO: Reinforcement Learning on Image Observations

  We trained PPO and DQN agents to play Super Mario Bros directly from raw pixels (stacked grayscale frames with the full 12-action control space). Our best agent achieved a completion rate of 96.32% on Level 1-1 using PPO and a top-k pseudo-greedy inference strategy. This performance is compared against other PPO and DQN agents. We investigated how architecture and preprocessing influences performance and evaluated transfer performance from Level 1-1 to Level 1-2.

• 

  A Comparative Study of Trading Using Deep Q-Learning and Double Deep Q-Learning in Niche Markets

  We investigated whether reinforcement-learning trading agents can perform well in niche markets by modeling trading as an MDP and comparing DQN vs Double DQN across three different action-space designs. Using data of the First North Growth Market and Novotek (with OMX30 as context), we evaluated performance with ROI and Sortino ratio across multiple runs for two assets. We showed how action-space choice influences stability and trading behavior, and that Double DQN generally delivers more robust risk-adjusted performance than standard DQN.

• 

  Predicting Hospital Readmission from Clinical Discharge Summaries

  We built a machine-learning pipeline to predict hospital readmission risk using only free-text discharge summaries from the MIMIC-IV dataset. We compared multiple text representations like keyword features, ClinicalBERT embeddings, and LLM-generated summary embeddings across several classifiers (Logistic Regression, tree models, XGBoost, and a neural net). Our best model (a Random Forest on LLM-based embeddings) achieved roughly F1 ≈ 0.79 and Accuracy ≈ 0.74, showing that clinical notes contain a signal for readmission prediction.

Other projects

• 

  Learning ROS 2: Swarm Controller

  I built a small ROS 2 drone-show simulation in Gazebo, where users can spawn any number of drones and command them into a scalable line or circle formation. A Python interface node uses an action client to send formation settings to a C++ controller node, which coordinates the drone nodes via publisher/subscriber communication. Live feedback of the progress is sent back to the interface.