hanyang9/UMR

What it solves

UMR addresses the challenge of motion retargeting for humanoid robots, specifically the difficulty of transferring complex human motions (from various sources like SMPL-X, BVH, or specialized datasets) to robots with different physical structures without requiring manual, joint-by-joint mapping between the human and the robot.

How it works

UMR uses the exterior body surface as a shared interface. The process happens in two stages:

1. Point Cloud Correspondence Learning: The system learns an ordered mapping between the surface of the source (human) and the target (robot) while they are in aligned canonical poses.
2. Correspondence-Guided Retargeting: The robot's motion is optimized based on the learned surface matches, taking into account positions, orientations, contacts, and kinematic constraints.

Once a correspondence is learned for a specific source template and robot, it can be reused for any motion using that same template.

Who it’s for

This tool is designed for robotics researchers and developers working with humanoid robots who need to convert human motion capture data or synthetic animations into executable robot trajectories.

Highlights

• Universal Source Support: Compatible with a wide range of motion sources including BONES-SEED, GRAIL, OmniContact, LAFAN1/SMPL-X, OMOMO, and standard FBX/BVH files.
• No Manual Mapping: Eliminates the need for manual human-to-robot joint mapping.
• Interaction Handling: Supports human-scene interaction (HSI) and human-object interaction (HOI).
• Batch Processing: Includes a pipeline for large-scale retargeting with bidirectional warm starts and dynamic programming to handle singularities.