Deployment of imitation-learning policies (ACT, SmolVLA, π0.5) on a SO-101 robotic arm: collected teleoperated demonstrations and fine-tuned policies that execute manipulation tasks on the arm. Full evaluation, quantified results, and failure analysis on the Robot Learning page.
Built an offline RL pipeline (TD3+BC) for vision-based manipulation on the WidowX arm in simulation, on the pick-and-place experimental setup from the COG paper. Replaced the CNN visual encoder with a pretrained ResNet-18 backbone, achieving faster and more stable convergence than a CQL baseline within 200 epochs. (code)
Fine-tuned a Vision Transformer with LaCLIP embeddings on the MMSD 2.0 (image–text pairs) dataset, taking advantage of text-augmented pre-training and to test whether the richer text supervision improves cross-modal reasoning. (code)