The project aimed to train a Reinforcement Learning algorithm for functional grasping and repositioning of a cordless drill. The robot used for this task comprised a Franka FR3 and an Agile Hand (five fingers). I am curremtly working on the integrtion of demonstrations to reduce the training time / reduce the need for reward engineering.
Key methods included a review of state-of-the-art algorithms for reinforcement learning and reinforcement learning from demonstrations applied to grasping and dexterous in-hand manipulation. A custom Reinforcement Learning environment was integrated into Isaac Sim to utilize massive GPU acceleration, running tasks at approximately 60,000 fps. A fine-tuned dense reward function was developed to enable natural motion of the expert policy, and a behavioral cloning pretrainer was created to bootstrap the policy.