webpage | abstract

We study perceptive root-guided humanoid–object interaction. We retarget optical motion-capture data together with object trajectories, and recover object trajectories from inertial motion-capture data using Blender animation. Building on the BeyondMimic framework, we track box-transport motions while removing reference-motion observations, using the robot root trajectory and contact information as control interfaces so the policy learns their relationship with object positions and distinguishes different phases of box transport. We deploy LIO, FoundationPose, and the motion-control stack on the G1's onboard AGX Orin, achieving fully onboard operation without external compute or motion capture. Root-trajectory planning enables arbitrary A-to-B continuous box transport, recovery after drops, and obstacle avoidance.

webpage | abstract

We use xy linear velocity, yaw angular velocity, and contact information as humanoid–object interaction (HOI) control commands. A teacher–student framework distills a privileged-information policy into an ego-view depth-map policy; the student estimates object states from depth maps and transfers to real hardware for diverse object interactions.

webpage | abstract

We propose a two-stage generalized mimic training pipeline: first training a mimic policy with reference-motion priors, then generalizing over ball positions and combining soccer-task rewards to train shooting policies for out-of-distribution static and moving balls. The approach achieves robust shooting on real hardware using only proprioception.

webpage | abstract

We develop a diffusion model that generates keyframes of human interaction motions and object trajectories from initial/final poses and object-path waypoints. An SMPL humanoid agent is trained in Isaac Gym to track target joints and object trajectories, achieving physically constrained dynamic object interaction. Generated motion sequences serve as policy tracking targets, forming a closed loop via an autoregressive generation structure to complete user-specified object interaction tasks.

webpage | abstract

We combine anchor representations with multi-resolution hash tables for compact Gaussian features, reducing peak memory and storage on large scenes. We derive an analytical solution for the low-pass filtering parameter under focal-length changes from stereo geometry, replacing empirical tuning and improving interpretability. During gradient-based 3DGS densification we preserve the sign of luminance gradients to avoid incorrect densification caused by insufficient brightness.