I worked on a humanoid robot trained using reinforcement learning to do
tasks such as pick and place.
? Perception: autoencoders, contrastive learning, 3D object pose
estimation. Robot achieved >90% success rate.
? Sensor fusion: made the system robust to losing the data from 1-2
out of 4 different sensors.
? Dynamics: created a model of the robot?s dynamics, which improved the
success rate in simulation by 5%.
? Behavioral cloning: trained policies for different tasks with success rates
above 90%.
? Engineering: made the real robot 3.3 times faster.