Amazon FAR Open Sources Holosoma to Unify Humanoid Simulation and Training

Amazon’s Frontier AI & Robotics (FAR) team has open-sourced Holosoma, a comprehensive software framework designed to streamline the complex pipeline of training and deploying humanoid robots.

Announced today by Amazon FAR researcher (and incoming UT Austin professor) Carlo Sferrazza, the release aims to solve what he describes as the "full-stack problem" of sim-to-real learning. By unifying fragmented simulation backends into a single codebase, Amazon is attempting to lower the barrier to entry for researchers working on next-generation bipedal locomotion and manipulation.

The release marks another significant contribution to the open science community from Amazon, following closely on the heels of their OmniRetarget data generation engine and the ResMimic manipulation framework.

Solving the "Back-End" Fragmentation

One of the persistent headaches in robotics reinforcement learning (RL) is the disconnect between various simulation environments. Researchers often train policies in high-throughput environments like IsaacGym, but validate them in physically accurate simulators like MuJoCo, before finally attempting to deploy them on real hardware.

Holosoma—derived from the Greek for "whole-body"—addresses this by supporting multiple simulation backends within a single training codebase. According to the repository documentation, the framework supports:

• Training: IsaacGym, IsaacSim, and MJWarp (MuJoCo Warp).
• Inference: MuJoCo (for simulation) and a shared pipeline for real-world deployment.

"With Holosoma, we unify the simulation landscape," Sferrazza wrote in a statement on X. "IsaacGym, IsaacSim, and MJWarp backends are all supported in a single training codebase."

The framework also integrates OmniRetarget, Amazon's previously released tool for converting human motion capture data into robot-feasible movements. This confirms that Amazon is building an interconnected ecosystem of research tools rather than isolated one-off projects.

Under the Hood: Algorithms and Accessibility

The framework is designed to be agnostic regarding the specific robot hardware, though it currently provides out-of-the-box support for the Unitree G1 and Booster T1 humanoids. This focus on the Unitree G1 aligns with industry trends, as the accessible price point of the G1 has made it a standard platform for research, also utilized in Amazon's ResMimic research.

Technically, Holosoma supports both locomotion (velocity tracking) and whole-body tracking tasks. It implements efficient RL algorithms including PPO (Proximal Policy Optimization) and FastSAC (Soft Actor-Critic), with native multi-GPU support to accelerate training times.

Notably, the release places a heavy emphasis on the "deployment" phase. A major bottleneck in academic research is often the "sim-to-real" gap—where code that works in a simulation fails on a physical machine due to latency or software incompatibility. Holosoma includes a unified inference stack that allows researchers to run the exact same code across simulation backends and on real-world robots, logging data directly to Weights & Biases (wandb) for analysis.

The Strategic Context

While Holosoma is an academic contribution, released under the permissive Apache-2.0 license, it fits into a clearer picture of Amazon's industrial ambitions.

Recent reports indicate Amazon has an internal goal of automating 75% of its operations over the coming years. Achieving this will require robots that can navigate unstructured environments and manipulate diverse objects—problems that traditional, hard-coded automation cannot solve.

By open-sourcing the tools required to train these "generalist" brains, Amazon accelerates the broader field of humanoid research while establishing its own methodologies as industry standards. The team behind Holosoma includes notable figures in the field, including Pieter Abbeel and others associated with the FAR team.

For Sferrazza, the release serves as a capstone to his time at Amazon FAR before he transitions to an assistant professorship at UT Austin, where he plans to continue research into humanoid robotics and dexterous manipulation.

The full codebase is available now on GitHub.