Skip to content

train_motion_tracking.md

Latest commit

 

History

History
141 lines (98 loc) · 5.65 KB

train_motion_tracking.md

File metadata and controls

141 lines (98 loc) · 5.65 KB
Raw

Train the Motion Tracking Model

After completing motion retargeting, you can train a motion tracking model with HoloMotion using the following process.

Overall Workflow:

flowchart LR
A[Motion Retargeting] --> B[HDF5 Database]
B --> C[Training Config]
C --> D[Training Entry]
D --> E[Distributed PPO Training]

classDef dashed stroke-dasharray: 5 5, rx:10, ry:10, fill:#c9d9f5
classDef normal fill:#c9d9f5, rx:10, ry:10
class A dashed
class B,C,D,E normal
Loading

1. Train v1.4.0 from scratch

The public v1.4.0 from-scratch path remains the baseline training entry. The training entry point is holomotion/src/training/train.py, which uses the training config to start distributed training across multiple GPUs.

1.1 Prepare the Training Config

Use holomotion/config/training/motion_tracking/motion_tracking_v1_4_0.yaml as the v1.4.0 template. Key configuration groups to modify are located under holomotion/config/:

  • /algo: Algorithm settings (PPO) and network configurations
  • /robot: Robot-specific config including DOF, body links, and control parameters
  • /env: Environment settings including motion sampling and curriculum learning
  • /env/observations: Observation dimensions, noise, and scaling for the policy
  • /env/rewards: Reward function definitions
  • /env/domain_randomization: Domain randomization settings (start with NO_domain_rand)
  • /env/terrain: Terrain configuration
  • /modules: The policy network modules definitions
# @package _global_

defaults:
  - /training: train_base
  - /algo: ppo_tf
  - /robot: unitree/G1/29dof/29dof_training_isaaclab
  - /env: motion_tracking
  - /env/terminations: termination_motion_tracking
  - /env/observations: motion_tracking/obs_motion_tracking_tf-moe
  - /env/rewards: motion_tracking/rew_motion_tracking
  - /env/domain_randomization: domain_rand_medium
  - /env/terrain: isaaclab_rough
  - /modules: motion_tracking/motion_tracking_tf-moe
  - _self_

project_name: HoloMotionMotrackV1.4

train_hdf5_roots:
  - { root: data/h5v2_datasets/AMASS_test, ratio: 1.0 }

Replace the example HDF5 root with the robot HDF5 output generated by HoloRetarget. Multiple roots can be assigned different sampling ratios.

1.2 Train your Policy

The training script defaults to motion_tracking_v1_4_0.yaml. Set CONFIG_NAME to select another config and NUM_ENVS to override the number of environments. Verify that train_hdf5_roots in the selected config points to the robot HDF5 dataset.

Start training by running:

bash holomotion/scripts/training/train_motion_tracking.sh

# Example override
NUM_ENVS=2048 CONFIG_NAME=motion_tracking_v1_4_0 \
  bash holomotion/scripts/training/train_motion_tracking.sh

Note that IsaacLab relies on internet connections to pull assets from Nvidia's cloud storage. If you encountered stuck at scene creation, it is very likely that you can't access the cloud-hosted assets. Turn on your proxy and try again can solve the issue.

2. Fine-tune v1.4.1 for modular G1 morphologies

v1.4.1 is a fine-tuned continuation of the published v1.4.0 model_14000 checkpoint. It is not a separate from-scratch recipe. The policy architecture, 29-DoF action order, and deployment input/output contract remain compatible with v1.4.0.

Initialize the public asset submodule and generate the 66 deterministic G1 variants:

git submodule update --init --recursive thirdparties/HoloMotion_assets

python assets/robots/unitree/G1/modular/g1_urdf_composer.py generate-all \
  --output-dir assets/robots/unitree/G1/modular/generated/urdf

Download the complete public v1.4.0 checkpoint package used as the fine-tuning base:

python holomotion/scripts/training/download_v1_4_checkpoint.py \
  HorizonRobotics/HoloMotion_models

Then start universal morphology fine-tuning explicitly:

CONFIG_NAME=motion_tracking_v1_4_1_universal_finetune \
  bash holomotion/scripts/training/finetune_motion_tracking_v1_4_0.sh

HOLOMOTION_FINETUNE_CHECKPOINT can override the default checkpoints/holomotion_v1.4/model_14000.pt path. The complete checkpoint package is required because fine-tuning restores actor and critic weights and normalization buffers while starting a fresh optimizer state.

Training Tips

How to use less GPU ?

Training requires significant GPU memory. Reduce NUM_ENVS if GPU memory is limited. This reduces both rollout and PPO memory consumption, but may make policy optimization less stable.

How to start multiple training session ?

In cases where you would like to start multiple training sessions, you should explicitly add the --main_process_port=port_number option in the training entry bash script to avoid port conflict of the accelerate backend. And this port_number can not be 0 .

If you would like to run training on a specific GPU, just modify the GPU id in the export CUDA_VISIBLE_DEVICES="X" statement.

How to set the save/log intervals ?

You may want to have more or less frequent logging and model dumping intervals. You can alter these intervals by adding the following options:

  • algo.config.save_interval=X : The checkpoint will be saved every X learning iterations.
  • algo.config.log_interval=Y: The logging information will be displayed every Y learning iterations.

Where is the checkpoint dumped ?

By default, checkpoints are written to logs/<project_name>/<timestamp>-<experiment_name>/. Change base_dir, project_name, or experiment_name to customize this path.

How to resume training from a checkpoint ?

To resume training, pass the checkpoint path as a Hydra override, for example: checkpoint=logs/<project_name>/<timestamp>-<experiment_name>/model_X.pt.