← Projects
Robust vision-language-action models
Bachelor thesis: can reinforcement learning improve a VLA policy after fine-tuning?
- Years
- 2026
- Status
- Finished
- Field
- research
About
Joint bachelor thesis at DTU Compute with Christian Amtoft Nickelsen and Celina Laungaard, graded 12. We measured how sensitive SmolVLA-0.45B is to visual and language perturbations on the LIBERO benchmark, then tested three sparse-reward RL post-training methods against a supervised baseline. Rewording an instruction without changing its meaning dropped closed-loop success close to zero on several tasks. The RL results tied with the baseline rather than beating it, and the calibrated evaluation protocol we built to establish that is the part I would defend hardest.
Built with
- PyTorch
- MuJoCo
- LIBERO
- SmolVLA
- Reinforcement learning