How horizonrobotics/holomotion works
HoloMotion is a humanoid-motion control system spanning motion-data preparation, policy learning, simulation evaluation, and robot deployment. The documented custom-policy workflow begins with HoloSMPL conversion, proceeds through robot retargeting and training, then evaluation; deployment also supports offline clips and live teleoperation without retraining. HoloSMPL provides a shared canonical motion representation and conversion, validation, packing, and visualization capabilities. The graph preserves these major branches. Several training and deployment links are documented rather than verified from source excerpts, so those edges cite the README; unsampled implementation details are not inferred.
Main entry points: cli.py HoloSMPL CLI · pack_formal_h5.py H5 packing · convert_dataset.py Canonical conversion · holomotion_preprocess.py Robot retargeting
Components 28
Motion data
- HoloSMPL CLI
holosmpl/cli.py - Source registry
holosmpl/sources.py - Source converters
holosmpl/converters - Canonical conversion
holosmpl/workflows/convert_dataset.py - Formal conversion
holosmpl/workflows/convert_formal.py - Distributed H5 production
holosmpl/distributed_holosmpl_h5.py - H5 packing
holosmpl/workflows/pack_formal_h5.py
Retargeting
- Robot retargeting
holomotion/src/motion_retargeting/holomotion_preprocess.py - HoloRetarget engine
holoretarget/_engine_impl.py
Learning and evaluation
- Robot H5 data
holomotion/src/training/data_production/robot_h5.py - H5 data loader
holomotion/src/training/h5_dataloader.py - Policy training
holomotion/src/training/train.py - PPO algorithm
holomotion/src/algo/ppo.py - Policy networks
holomotion/src/modules/network_modules.py - Simulation evaluation
holomotion/src/evaluation/eval_mujoco_sim2sim.py - Motion tracking environment
holomotion/src/env/motion_tracking.py
Robot runtime
- Offline deployment
deployment/unitree_g1_ros2_29dof/src/humanoid_policy/motion_clip_library.py - Policy runtime
deployment/unitree_g1_ros2_29dof/src/humanoid_policy/policy_runtime.py - ROS 2 policy node
deployment/unitree_g1_ros2_29dof/src/humanoid_policy/policy_node_29dof.py - Live teleoperation
deployment/holomotion_teleop/holomotion_teleop_node.py
Shared data formats
- Data validation
holosmpl/core/validation - Canonical schema
holosmpl/core/schema/canonical.py - Motion resampling
holosmpl/core/processing/resample.py - Formal H5 writer
holosmpl/core/writers/formal_h5.py - Motion visualization
holosmpl/visualization/video_render.py
Other components
- Motion user
- Unitree G1
- Live motion source
Connections 31
- Motion user → HoloSMPL CLI: converts motion
README.md - HoloSMPL CLI → Source registry: selects source
holosmpl/cli.py - Source registry → Source converters: dispatches to
holosmpl/sources.py - HoloSMPL CLI → Canonical conversion: runs conversion
holosmpl/cli.py - Canonical conversion → Canonical schema: writes canonical data
holosmpl/workflows/convert_dataset.py - Canonical conversion → Data validation: validates output
holosmpl/workflows/convert_dataset.py - HoloSMPL CLI → Formal conversion: runs conversion
holosmpl/cli.py - Formal conversion → Data validation: validates formal data
holosmpl/workflows/convert_formal.py - HoloSMPL CLI → H5 packing: packs formal data
holosmpl/cli.py - H5 packing → Data validation: checks H5 data
holosmpl/workflows/pack_formal_h5.py - H5 packing → Formal H5 writer: writes H5
holosmpl/workflows/pack_formal_h5.py - HoloSMPL CLI → Distributed H5 production: runs distributed production
README.md - Distributed H5 production → H5 packing: combines H5 shards
holosmpl/distributed_holosmpl_h5.py - HoloSMPL CLI → Motion visualization: renders previews
holosmpl/cli.py - Motion user → Robot retargeting: retargets motion
README.md - Robot retargeting → HoloRetarget engine: uses engine
- Robot retargeting → Robot H5 data: produces robot data
README.md - Robot H5 data → H5 data loader: provides training data
holomotion/src/training/h5_dataloader.py - Motion user → Policy training: trains policy
README.md - H5 data loader → Policy training: supplies batches
README.md - Policy training → PPO algorithm: optimizes with PPO
- PPO algorithm → Policy networks: uses networks
holomotion/src/algo/ppo_tf.py - Motion user → Simulation evaluation: evaluates policy
README.md - Simulation evaluation → Motion tracking environment: evaluates tracking
- Motion user → Offline deployment: replays clips
README.md - Offline deployment → Policy runtime: provides reference motion
- Live motion source → Live teleoperation: streams motion
README.md - Live teleoperation → Policy runtime: feeds live references
deployment/holomotion_teleop/holomotion_teleop_node.py - Policy runtime → ROS 2 policy node: runs policy
- ROS 2 policy node → Unitree G1: controls robot
README.md - Live teleoperation → Unitree G1: supports teleoperation
README.md