Overview

HALO is the code release accompanying the paper Perceptive Humanoid Motion from Simulation. It targets researchers who want to reproduce or build on a pipeline that generates humanoid locomotion data in simulation, trains a motion generator and a motion-tracking policy, and deploys the policy on a Unitree G1, in MuJoCo simulation or on the real robot. A language orchestrator turns a request plus an egocentric image into a plan of subtasks.

The repository is organized by stage: datagen/ builds synthetic training data from Infinigen scenes, Kimodo motion and Isaac/MuJoCo rendering; generator/ covers motion-generator training and open-loop evaluation; tracker/ holds the G1 tracking policy with Isaac Lab training and Python/C++ deployment; orchestrator/ runs the planning model; system/ ties the nodes together into closed-loop sim or real-robot runs. Robot and scene assets sit in assets/.

Getting started

Start from a specific subproject rather than expecting a single top-level install. The README routes each workflow to its own instructions:

1. Generate training data following the data generation guide.
2. Train the generator with the generator README; train the tracker through the tracker training guide, which covers Isaac Lab and the three phases: privileged experts (PreRL), multi-teacher distillation, and multi-critic fine-tuning (PostRL).
3. Deploy the tracker alone via the Python deployment guide or the real-time C++ guide.
4. Run the closed loop with run_sim.py in MuJoCo per the simulation section, or on a G1 with run_real.py per the real-robot section.

Setup commands, environments and prerequisites live in those per-folder READMEs; the top-level README does not define a single combined installation.

Constraints

• The pipeline assumes substantial simulation tooling (Isaac Lab, MuJoCo) and a Unitree G1 for the real-robot path.
• No pretrained tracker weights are shipped; the proprio and depth tracker students must be trained with tracker/training.
• The repository notes open TODOs around release scope, the author list and licensing; third-party assets such as the Unitree G1 description keep their original licenses. Check the LICENSE file before use; GitHub reports the license as unrecognized (NOASSERTION).

HotRepo's take

The project appeared on HotRepo's new-project list for the Python board on 2026-11-06, reaching 108 stars (up 6 within the recorded growth window) and 8 forks, with no GitHub daily rank recorded. The material is a multi-stage research release with per-module documentation, which makes it a reference for reproducing the HALO pipeline rather than a drop-in package; the open TODOs around weights and licensing are the practical limits to weigh first.