Undergraduate Thesis: Human-in-the-Loop Learning for Humanoids Robots (Presented at Humanoid Conference Workshop)

I completed my Turing Honors Thesis!

Abstract - Recent advances are demonstrating humanoid robots’ immense potential to operate effectively in our human- centered world. Deep imitation learning offers a promising approach for training autonomous humanoids through human demonstrations via teleoperation. However, due to the robots’ high degree of freedom and need for stabilized robot dynamics, a large number of task demonstrations is necessary to train effec- tive models. In order to reduce the human workload in the pre- training process, we propose a system that incorporates human- in-the-loop interactive learning during humanoid deployment to enhance learning efficiency. In simulated experiments on the DRACO 3 humanoid robot, we showed that our system can attain over 80% task success with three times fewer demonstrations than when learning from scratch. Our study illustrates the effectiveness of human-robot collaboration in improving humanoid robot skills and autonomy.