DEFT: Dexterous Fine-Tuning for Real-World Hand Policies

The longstanding goal of robot learning is to build robust agents that can perform long-horizon tasks autonomously. This could for example mean a self-improving robot that can build furniture or an agent that can cook for us. A key aspect of most tasks that humans would like to perform is that they require complex motions that are often only achievable by hands, such as hammering a nail or using a screwdriver. Therefore, we investigate dexterous manipulation and its challenges in the real world. A key challenge in deploying policies in the real world, especially with robotic hands, is that there exist many failure modes. Controlling a dexterous hand is much harder than end-effectors due to larger action spaces and complex dynamics. To address this, one option is to improve directly in the real world via practice. Traditionally, reinforcement learning (RL) and imitation learning (IL) techniques have been used to deploy hands-on tasks such as in-hand rotation or grasping.