Human Perception of Whole-Body Robot Motion

Motivation

Background

Offline human-to-robot motion retargeting converts human demonstrations into robot movements, typically with the goal of reproducing task-space behavior, such as end-effector trajectories, so that the robot achieves the same manipulation effect on the environment.
Gap

Such retargeted trajectories can therefore serve as offline robot demonstrations for downstream behavior cloning. However, task-space tracking accuracy alone does not establish whether people trust the robot, feel comfortable around it, or perceive its motion as natural and understandable.
This project

This project compares user perceptions of robot motions generated by WARP and Mink-based baselines across several HRI dimensions. A diverse set of tasks from both simulation and real-world recordings will be used as study stimuli.

Retargeting Methods

All methods follow the same human demonstration with the robot’s hands. They differ in how the rest of the body (torso, elbows, overall posture) is resolved, and that difference is what participants see.

Left to right: reference human motion, WARP, and the Mink baseline, first tracking the end effector only (MINK-EF) and then with torso and elbow targets added (MINK-TE). Video from the WARP project.

Closed-form retargeting

WARP

Matches the palm pose exactly as a hard constraint, then uses the remaining freedom to preserve the human’s shoulder–elbow–wrist posture. Joint angles follow analytically, giving one unique solution per frame.

Optimization-based IK

Mink-based baselines

Weighted inverse kinematics built on the Mink library: palm-pose tracking with optional torso-orientation and elbow-swivel objectives. MINK-EF tracks the end effector only; MINK-TE adds the torso and elbow terms.

Robot Platform

The robot in our study is the Rainbow Robotics RB-Y1, a wheeled humanoid whose kinematics mirror the human upper body. Like a person, it is kinematically redundant: the same hand pose can be reached with many different arm and torso postures. Each retargeting method has to choose one, and that choice is what observers see.

The RB-Y1 wheeled humanoid in a kitchen: two arms with five-fingered hands, a head carrying Project Aria glasses, a segmented torso, and a black mobile base. — The RB-Y1 platform used in the WARP real-world experiments.

Project Aria Gen 2 glassesHead-mounted egocentric sensing
2-DoF neckOrients the head and its sensors
12-DoF XHandsFive-fingered dexterous hands
Two 7-DoF armsRedundant: one hand pose admits many elbow poses
6-DoF torsoRepositions the upper body without moving the base
Holonomic mobile baseDrives in any direction across the floor