Sim-to-real learning for humanoid robots is a full-stack problem. Today, Amazon FAR is releasing a full-stack solution: Holosoma. To accelerate research, we are open-sourcing a complete codebase covering multiple simulation backends, training, retargeting, and real-world inference.

Simulation frameworks such as IsaacGym, IsaacLab, MuJoCo Playground, mjlab were all great inspiration to us! With Holosoma, we unify the simulation landscape: IsaacGym, IsaacSim, and MJWarp backends are all supported in a single training codebase. We also support multiple robots and RL algorithms: PPO and FastSAC for very efficient learning — and offer native multi-GPU support!

A major barrier in humanoid research is the lack of fully open, low-latency inference pipelines. Therefore, we release our inference stack, which allows you to run the exact same code across simulation backends and on real-world robots.

We support both velocity tracking and whole-body tracking tasks. For whole-body tracking, we address retargeting with an OmniRetarget open-source implementation and re-implement simple BeyondMimic-style tracking. OmniRetarget: BeyondMimic:

A feature I find really helpful is video logging to wandb across all simulation backends. So much easier to sweep hyperparameters and browse through the results when you get to actually see the policy behavior on the screen. We also log ONNX files as we train, and our inference pipeline supports loading them directly through wandb, which makes sim-to-real iterations astonishingly fast!

We focused heavily on simplicity and modularity. Reward engineering is kept to a minimum, and we support manager-based environments (similar to Isaac Lab and mjlab) for ease of adaptability. In fact, Holosoma was built to be extended! We are releasing an example extension showing how to adapt the relevant classes to support training and inference on a Unitree Go2 quadruped robot:

Our goal is to lower the barrier to entry for humanoid research. By providing a broad infrastructure backbone, we hope to empower researchers to focus on novel algorithms and behaviors. This was all built by amazing people at Amazon FAR: @pabbeel, @jychen_1729, @rocky_duan, @alescontrela, @mr3gandhi, Samuel Gundry, @x_h_ucb, @akanazawa, @tomaszwlewicki, @jiaman01, @ckarenliu, Clay Rosenthal, @younggyoseo, @GuanyaShi, Linda Shih, Jonathan Tseng, @zhenkirito123, @lujieyang98, @brenthyi, @Yuanhang__Zhang

All of this is open-source, and you can get started now at: Example extension for Go2 robot: We can't wait to see what the robotics community builds with Holosoma!

Great work! Congrats to all the dear friends loll

Keep at it, buddy. We need more people in this robotics age. No others can beat the usefulness of a good humanoid robotics.

🙏

Nice work team at far, look forward to trying this

Amazing work!

Nice release!

So excited to try it!

U are an ass whooolll’mother fucker.

Open-sourcing everything from simulation to real-world inference will accelerate adoption fast.

This is a huge contribution. The full-stack sim-to-real problem is *the* bottleneck. Open-sourcing the entire pipeline is a massive accelerator for the field. Can't wait to dig in.

Sim-to-real has always struggled because everyone tackled pieces instead of the full pipeline. Seeing Amazon FAR ship Holosoma as a complete, open-source stack finally closes that gap. A unified codebase for simulation, training, retargeting, and real-world inference is exactly what the field needs. This is the kind of foundation that accelerates humanoid progress instead of fragmenting it.

Unifying IsaacGym, IsaacSim, and MJWarp into a single training codebase fundamentally changes sim-to-real. This enables physics-engine-level domain randomization, an aggressive, data-driven strategy to minimize the reality gap for humanoid RL research at scale.

Awesome release. Open sourcing the full stack is the real unlock for applied work in this space. So much of the challenge is in the infra. Great to see.

🧠✨😎👍

Idiot, I'm the one who should be dancing. And this fucking robot should be working in a factory, in a field, in any manufacturing or service sector 24/7. Useless crap, fuck you.

cool that they are opensourcing that. gotta be a game changer for humanoid research. wonder how it stacks up against existing tools?