Holosoma Updates Announcements

[samuelgundry]

Hi All,

We have pushed from our internal development repository. Please see the Upgrades section below for required dependency updates.

Note: We're working on transitioning to GitHub as our primary repository once we have publicly visible CI for PRs and training dashboards in place. We'll then post PRs to Github rather than push commits from our internal repository. We will also then be ready to merge community PRs using standard practices.

Thank you for the feedback. We're excited to share and collaborate with this community.

Upgrades

IsaacSim 5.1 / IsaacLab 2.3

• You will need to reinstall: ./scripts/reset_isaacsim.sh && ./scripts/setup_isaacsim.sh

Tyro 1.0.0 and Unitree SDK 0.1.2

• Update each of your Training Conda environments (hsmujoco, hsgym, and hssim)
  • For example: source scripts/source_mujoco_setup.sh && pip install -e .src/holosoma[unitree]
• Inference and retargeting environments do not require updates at this time

Updates and Improvements

Whole-Body Tracking

• Added scripts for running complete pipeline for whole-body tracking, including retargeting, data conversion, and training
• Added replay script for wbt and update root height to 0.76
• Fixed WBT inference: removed yaw_offset in motion_clip torso_ori
• Fixed several small bugs in WBT

Retargeting

• Added AMASS SMPLX data processing code and SMPLX support for retargeting
• Added instructions for custom human motion data format
• Added instructions for custom robot type
• Updated viser visualization with synchronized frame number display
• Removed hard-coded values and cleaned up unused files/functions

Inference and Sim2sim

• Moved sim2sim Unitree SDK to C++ bindings, removes unitree_sdk_python dependency

Terrain

• Exposed terrain spawn parameters to config with opt-in height querying for resetting/randomization on rough terrain

Infrastructure

• Added Dockerfile

Documentation

• Improved README for headless video recording
• Added note that replay buffer size is per env
• Added demo_scripts/ reference in README

[sutyum]

Moved sim2sim Unitree SDK to C++ bindings, removes unitree_sdk_python dependency

I have a somewhat stupid question for the authors: isn't the SDK just a wrapper over their DDS API, so i would expect the language choice to have relatively low effect on performance. Was an actual problem faced? Was an actual improvement achieved? (and measured)