Scene-Aware Humanoid Locomotion and Policy Inference

Moving Through Clutter plus an ongoing ROS 2/C++ inference prototype for humanoid policies

Moving Through Clutter

Real homes and workplaces are dense, three-dimensional, and geometrically constrained, while many humanoid locomotion systems are demonstrated on open, flat terrain. Moving Through Clutter (MTC) (Wang et al., 2026) is a virtual-reality framework for collecting and evaluating embodiment-consistent, scene-aware humanoid motion in procedurally generated clutter.

The framework captures whole-body human motion in VR, automatically retargets it to a humanoid model, and measures stability, collision safety, and scene clutter. The first dataset contains 348 trajectories across 145 diverse 3D scenes. I am third author, with Beichen Wang, Yuanjie Lu, Liuchuan Yu, and Xuesu Xiao.

Humanoid policy-inference prototype

Alongside the research, I am prototyping a deployment path for humanoid motion policies in C++ and ROS 2. The public motion-tracking controller repository adds:

  • paired ONNX residual and base-policy inference;
  • metadata-driven observation assembly and normalization;
  • temporal history buffering and policy diagnostics; and
  • an Isaac Lab-to-MuJoCo parity workflow before longer simulation or hardware experiments.

This is an ongoing prototype, not a claim of validated physical deployment. Its validation gates remain explicit in the repository documentation.

Research direction

MTC’s controllable scene-clutter levels provide a natural task axis for future curriculum-learning research. Connecting automatic curricula to scene-aware whole-body control is an active direction rather than a reported result of the MTC preprint.

Resources

References

2026

  1. arXiv
    Moving Through Clutter: Scaling Data Collection and Benchmarking for 3D Scene-Aware Humanoid Locomotion via Virtual Reality
    Beichen Wang, Yuanjie Lu, Linji Wang, Liuchuan Yu, and Xuesu Xiao
    arXiv preprint arXiv:2603.05993, 2026