Robot-assisted dressing could benefit the lives of many people such as older adults and individuals with disabilities. Despite such potential, robot-assisted dressing remains a challenging task for robotics as it involves complex manipulation of deformable cloth in 3D space.  Many prior works aim to solve the robot-assisted dressing task, but they make certain assumptions such as a fixed garment and a fixed arm pose that limit their ability to generalize. In this work, we develop a robot-assisted dressing system that is able to dress different garments on people with diverse poses from partial point cloud observations, based on a learned policy.  We show that with proper design of the policy architecture and Q function, reinforcement learning (RL) can be used to learn effective policies with partial point cloud observations that work well for dressing diverse garments.  We further leverage policy distillation to combine multiple policies trained on different ranges of human arm poses into a single policy that works over a wide range of different arm poses. We conduct comprehensive real-world evaluations of our system with 510 dressing trials in a human study with 17 participants with different arm poses and dressed garments. Our system is able to dress 86% of the length of the participants' arms on average.

All videos are 4x real-time. All dressing trials are performed by a single policy using partial point cloud observations.

The baseline policy, which does not use policy distillation to generalize to diverse poses, has difficulty correctly turning at the eblow and dressing the upper arm in most trials.

In simulation, we leverage reinforcement learning and policy distillation to learn a single policy that is able to work with diverse arm pose ranges, body sizes & shapes, and 5 different garments. The gifs below show our policy's dressing performance in simulation. 

In simulation, we evaluate when the person does a one-time movement of the arm after we record the initial arm point cloud. The movement is quantified as the joint angle changes at shoulder and elbow.  We test three joint angle changes: lowering down the shoulder, lowering down the elbow, and bending inwards the elbow, which are three common unconscious movements we observe in our real-world human study. 

The figures below show how the upper arm dressed ratio changes when the joint angle change increases. The first plot shows how the upper arm dressed ratio change with respect to shoulder joint bending dowards, the second with respect to elbow joint bending downwards, and the third with respect to elbow joint bending inwards. Each colored line shows the performance of one garment.

We find our system to be robust to 8.6 degrees of change in joint angles (averaged across three types of joint angle changes and 5 garments) while maintaining 75% of the original upper arm dressed ratio.

With the same single-arm dressing assumptions (static arm pose and robot already grasping the garment), we show that our method generalizes to dual-arm dressing as well.  Our Dense Transformation policy can handle controlling of two robotic arms well by simply extracting actions corresponding to both of the robot gripper points. 

In simulation, we successfully train two indivisual policies to perform dual arm dressing of a hospital gown (left) and a cardigan (right) on a fixed pose.