At BMW, I work on Vision–Language–Action (VLA) models — intelligent systems that leverage visual input and natural language instructions to predict robotic actions.
My focus is on applying reinforcement learning to fine-tune these models for specific assembly tasks, building on the generalized knowledge of large pre-trained models. More specifically, I use demonstration-guided reinforcement learning via the SERL method[1] to enable rapid skill acquisition through online learning.
The project focuses on dexterous manipulation for assembly tasks. The approach builds on ideas introduced in "Imitation Bootstrapped Reinforcement Learning" by Hengyuan Hu et al.: it begins by learning a policy from expert demonstrations, which is then reused as a guiding prior to improve exploration efficiency and learning stability during reinforcement learning, in parallel with directly incorporating demonstrations into the RL pipeline.
Using this approach, the policy learned the assembly task in about one hour of real-world practice — with real-world rollouts and improvement, and without relying on simulation.
Learned assembly policy, deployed on the real robot.