Robotics5 min read

HoloMotion v1 Adds ROS2 Deploy Paths for Unitree G1

Horizon Robotics HoloMotion v1.0 documents ROS2 and ONNX deployment for whole body humanoid control on Unitree G1 class hardware.

Key points

  • Horizon Robotics released HoloMotion v1.0 on 5 November 2025.
  • HoloMotion is an open source stack for whole body humanoid control.
  • Humanoid pilots fail when sim policies break on contact rich whole body tasks.
Published November 3, 2026
HoloMotion v1 Adds ROS2 Deploy Paths for Unitree G1
Kindel Media / Pexels

What changed

Horizon Robotics released HoloMotion v1.0 on 5 November 2025. HoloMotion is an open source stack for whole body humanoid control. The README covers training, tests, and ROS2 deploy steps. The stated target hardware includes the Unitree G1.

The v1.0 drop follows a research preview. It adds a deploy module that exports ONNX models for real time control. Teams can move from sim training to on robot loops with fewer mystery steps.

Why it matters

Humanoid pilots fail when sim policies break on contact rich whole body tasks. Walking, reaching, and recovery all stress the same controller. A ROS2 deploy story bundled with training code lowers the glue work. That helps teams that already use Unitree style platforms.

For robotics product leads, this is a signal. Open whole body stacks are moving from paper demos toward repeat bring up steps. Investors should compare Horizon’s release cadence against public lab tests this month.

Safety remains non negotiable. The README warns that real robot tests carry injury risk. Treat v1.0 as engineering software, not as a consumer toy. Estops and joint limits still belong to your site rules, not to the repo alone.

Who is affected

Humanoid software teams feel this first. So do sim to real engineers and field test groups on G1 class hardware. Investors tracking Chinese robotics open source should map Horizon’s cadence against rival humanoid repos.

Systems integrators who sell ROS2 services get a new reference pipeline. They can compare it against in house controllers on the same robot. University labs gain a shared baseline for course projects and thesis work.

Insurance and site safety officers should review humanoid test plans even when code is open. Open source does not remove liability when a robot falls or hits a bystander.

What to do next

Clone the release tag and run the documented deploy dry run on your target robot. Do this before you merge policy code into production schedulers. Treat safety interlocks and estops as mandatory on every test day.

If you already use Isaac style simulators, test whether HoloMotion retargeting fits your motion library. You may need custom adapters for your gripper or head cameras. Log sim to real gap metrics so product teams see honest numbers.

Share failure videos inside your team, not only success clips. Whole body control debugs faster when everyone sees the same slip or stumble events.

What to watch

Watch the promised v1.1 hardening notes on the main branch. Watch whether third party G1 users publish latency numbers for ONNX inference on edge GPUs.

Track Open X Humanoid and other labs for competing sim to real stacks. TienKung Lab added its own real robot checks in November 2025. Compare license terms and hardware lock in before you bet a product road map on one stack.

Regulators in some regions may ask for risk docs when humanoids leave fenced labs. Keep test logs and incident notes even for early pilots.

Field service teams should track spare part lead times for G1 class legs and wrists. Control software means little if hardware waits on slow shipments.

Hiring managers can use HoloMotion as a skills test for ROS2 hires. Ask candidates to explain the deploy module and how they would add a new sensor frame. Answers reveal more than quiz trivia on link names.

Warehouse ops leads should keep humans out of the test cell until controllers pass a written safety review. Even slow walks can hurt if a policy runs untested on real hardware.

Media teams should label demo clips as research builds, not as shipping products. Clear labels reduce public hype that sales teams must later walk back.

Budget owners should count engineer weeks for sim tuning, not only GPU hours. Whole body stacks eat calendar time when policies fail on first contact.

Keep a pinned Docker or conda file next to the release tag you tested on hardware.

Image: GitHub

Photo: Kindel Media / Pexels

Sources

Related in Robotics

Continue reading

Boston Dynamics Atlas Trains Whole Body Fridge Lift in Simulation