Humanoid whole-body control is shifting from many specialized controllers toward unified policies trained on large motion corpora. This repository packages the pieces researchers need to train, evaluate, and deploy such policies end-to-end — from data collection and large-scale training to low-latency C++ inference and VR teleoperation.
What Sets It Apart
- GEAR‑SONIC foundation policy and released checkpoints — so what: provides a single policy architecture and released weights designed to generalize across diverse whole‑body behaviors (walking, crawling, manipulation), reducing the need to build per-skill controllers.
- Decoupled WBC (RL lower body + IK upper body) used in GR00T N1.5/N1.6 — so what: combines learned locomotion with deterministic upper‑body kinematics for stable teleoperation and manipulation.
- Full deploy/teleop stack (C++ inference, motor error monitoring, ZMQ v4, PICO VR teleop) — so what: moves beyond simulation by addressing low‑latency inference, robot monitoring, and real‑time human-to-robot motion transfer.
- Large-scale data & training pipeline (Bones‑SEED dataset, VLA collection tools, Hugging Face releases) — so what: enables foundation‑scale motion training and fine‑tuning, but expects substantial compute and robust data management.
Who it's for — and tradeoffs
Great fit if you need a research-to-robot pipeline for humanoid whole‑body control: teams that want pretrained foundation policies, VR teleoperation for data collection, or a C++ inference stack for real hardware. Look elsewhere if you need a minimal, dependency‑free demo — the repo relies on Isaac Lab, Git LFS, MuJoCo tooling, and substantial compute for training/finetuning (the project documents recommend large multi‑GPU setups for serious training). Also note the dual licensing: code under Apache‑2.0 but model weights under NVIDIA Open Model License, which carries additional usage terms.
Where it fits
Positioned between academic motion‑imitation codebases and full robot deployment stacks: it integrates large motion datasets and foundation policy checkpoints (Hugging Face links provided) with pragmatic deployment features so teams can iterate from imitation training to real‑time robot control.
