Research
I work on embodied AI: models that take an image and an instruction and produce something a robot can execute. My bachelor thesis measured how far reinforcement learning can push one of them past its supervised baseline. The answer was not very far, and the reasons are what I want to keep working on.
Interests
Vision-language-action models
Policies that map an image and an instruction straight to robot actions. I am interested in where that mapping is brittle and what the failures say about the representation underneath.
Reinforcement learning after supervised fine-tuning
Sparse-reward post-training on top of a strong imitation baseline. The open question is where the headroom actually is, since tasks near saturation give almost no learning signal.
World models and predictive representations
JEPA-style objectives that predict in latent space rather than pixels, and whether the embeddings they produce are dense enough to reward a policy with.
Multimodal robustness
What breaks a policy when the input shifts. Rewording an instruction should not change what a robot does; often it does.
Embedded and real-time systems
The other half of my background. Control loops on microcontrollers, and the distance between a gait that works in simulation and one that works on a real servo.
Output
- 2026 Preprint in preparation
Towards Robust Vision-Language-Action Models: Multimodal Diagnostics and Reinforcement Learning Post-Training
Rune Daugaard Harlyk, Christian Amtoft Nickelsen, Celina Laungaard
BSc thesis at DTU Compute, graded 12. We diagnosed the vision and language sensitivity of SmolVLA-0.45B on LIBERO and then tested whether sparse-reward RL post-training could beat the supervised baseline. Semantically neutral rewordings of an instruction drove closed-loop success close to zero on several tasks, which says the policy leans hard on the LIBERO template rather than on meaning. On the RL side we built a calibrated evaluation protocol — seeded initial states, MuJoCo 3.3.2, 100 episodes per task — under which the public checkpoint scores 79.9% on LIBERO Spatial. Our strongest multi-task result reached 80.5%, which is within noise of the baseline. The useful conclusion is that benchmark tasks already close to saturation give sparse-reward methods almost nothing to learn from.
Other work
Fooling Neural Networks: A Benchmark of Adversarial Attacks and Defenses
Group coursework benchmarking I-FGSM, PGD and Carlini-Wagner against MobileNet, ResNet and Swin on CIFAR-10 and ImageNet-20, including adversarially trained and architecturally robust variants. It sits next to the thesis because both are about what makes a learned model fragile.