Human-G1 Interaction

Supplementary Materials

1. SMPLX Mesh Projection

We project the SMPLX human mesh surface onto both the G1 robot mesh and SMPLX-fitted capsule primitives. This enables dense contact correspondence between human and robot body surfaces.

SMPLX to G1 Surface Projection

SMPLX to G1 surface projection
SMPLX mesh vertices projected onto the G1 robot mesh surface. Each color represents a different body part correspondence.

SMPLX Capsule Registration

SMPLX capsule registration
Capsule primitives fitted to SMPLX body parts.
Capsule assignment visualization
Body part assignment to capsule primitives.

SMPLX to Capsule Projection

SMPLX to capsule projection
SMPLX mesh projected onto capsule surfaces.
Projection on capsule
Projected contact points on capsule primitives.

2. Retargeting Process & Results

Our retargeting pipeline transfers human motion captured via SMPLX to the G1 humanoid robot, preserving contact semantics through a multi-stage optimization process.

Interactive Viewer

Retargeted motion sequences can be viewed interactively on our HuggingFace Space viewer .

Contact Refinement Results

Retarget comparison images and videos will be added here.

3. Policy Training

3.1 Architecture

The policy network takes as input the robot proprioception, retargeted reference motion, and contact state. It outputs joint position targets for the G1 robot.

Policy architecture diagram will be added here.
Input observations, network structure, output action space.

3.2 Reward Design

Reward components breakdown will be added here.
Motion tracking, contact reward, regularization terms.

3.3 Experiments & Results

Evaluation Videos

A011 — With proximity reward
A016 — With proximity reward
A000 — With proximity reward
A022 — With proximity reward