RCareWorld supports robots that are commonly used in physical and social caregiving, including Pepper, NAO, Stretch RE1, Kinova Gen3, Kinova Gen2, PR2, UR5, and Franka. For haptics sensing, users can get force and torque on the joints of these robots. For visual sensing, users can mount cameras on the wrist or head of these robots.

We use SMPL-X models to represent a human avatar’s shape and expression, and integrate it with an actuatable musculoskeletal backbone to model kinematics and dynamics. We use a behavior tree to model social behavior and individual preferences.

Using clinical data collected from real users with varying levels of spinal cord injury - both partial and complete, cerebral palsy, and hemiparetic stroke, we construct CareAvatars for six care-recipients.

CareHomes are virtual homes which accurately model caregiving homes with modifications for varying levels of accessibility and contain various assistive devices.

By consulting stakeholders, including occupational therapists, caregivers, and care-recipients, and following the guidance of Universal Design Manual, we present we present three levels of modifications:

• Level 1 - Usual: No special modifications
• Level 2 - Partially Barrier-Free: We modify some objects that are frequently used
• Level 3 - Completely Barrier-Free

We provide a set of six realistic robotic caregiving task scenes which can provide meaningful assistance to the modeled care-recipients with activities of daily living. Researchers can use these scenes to perform experiments for robotic caregiving, such as benchmarking reinforcement learning algorithms or testing planning and control algorithms.

Task: Care-recipient U6 is fed with a Kinova Gen 3 mounted on a wheelchair. The robot is required to move a soft and deformable food-item on the spoon near the care-recipient’s mouth and insert it into the mouth.

Task: Care-recipient U2 is lying on a hospital bed. A Kinova Gen 2 robot mounted on the bed holds a soft sponge in its gripper. The robot wipes the care- recipient’s arm with the sponge.

Baseline: Proximal Policy Optimization (PPO)

Task: We consider the scene where a human care- giver and a robot caregiver collaborate with each other to help care-recipient A4 put on shorts. The care- recipient is sitting on a wheelchair. The human caregiver lifts the care-recipient’s leg. The robot caregiver, Kinova Gen 3, mounted on the wheelchair pulls up the shorts from ankle to hip.

Baseline: Proximal Policy Optimization (PPO)

Task: When care-recipient U3 is transferred from their bed to a wheelchair using a hoyer sling, the limbs are usually not in the correct position (e.g. resting on the wheelchair handle) after the care-recipient is transferred from the hoyer sling. In this task, our human avatar sits in a wheelchair, and their limb is hanging down. The robot, Kinova Gen 3, mounted on the wheelchair holds a C-shaped tool to lift their upper limb, and put their arm on the wheelchair handle.

Baseline: Hindsight Experience Replay (HER) with Truncated Quantile Critics (TQC)

In this task, we want to demonstrate that our CareHomes assets can affect the skills that a robot learns in an environment. The Kinova Gen 3 arm mounted on the wheelchair adapts to different door mechanisms and opens them. We show a door knob for a level 1 home, a lever for a level 2 home, and a sliding door for a level 3 home.

Task i: Door with Knob in or Level 1 CareHome

Baseline: Hindsight Experience Replay (HER) with Soft Actor-Critic (SAC)

Task ii: Door with Lever Handle in Level 2 CareHome

Baseline: Hindsight Experience Replay (HER) with Soft Actor-Critic (SAC)

Task iii: Sliding Door in Level 3 CareHome

Baseline: Hindsight Experience Replay (HER) with Soft Actor-Critic (SAC)

Task: We consider a Stretch RE1 robot helping open a toilet cover in a bathroom. .

Baseline: Hindsight Experience Replay (HER) with Soft Actor-Critic (SAC)