Duke Humanoid v1

A machined research platform with substantial commissioning

The public bill-of-materials page distinguishes symmetric parts from mirrored left/right parts and exposes a spreadsheet, a CNC/DIY archive and full Fusion assemblies. Their availability is visible in the wiki; this review has not preserved or dimensionally validated those manufacturing downloads. The manufacturing guide specifies aluminum 6061, ISO 2768 medium tolerance and Ra3.2 roughness for outsourced CNC parts, alongside part-specific PLA, PLA-CF and TPU printing settings. [S007] [S008]

The electronics guide describes Motorevo PA100 motors with several reduction ratios, EtherCAT communication and Elmo Studio configuration. Its laboratory power arrangement uses four 48 V Meanwell supplies with current sharing and a capacitor. These are source-reported system choices, not an independently checked wiring design or a current purchasable kit. No complete delivered build budget is established here. [S009]

Resolve the setup filenames before following the examples

The hardware README reports Ubuntu 22.04 with CUDA 12.3, GCC 12, CMake, vcpkg and a Python 3.8 environment. The simulation route uses Isaac Gym and rl_games. Both pinned READMEs name setup/conda_env_py38.yaml, but that file is absent from the inspected archives. The actual environment files are setup/conda_env.yaml in hardware control and envs/setup/conda_env.yaml in simulation. Their presence does not establish that they are interchangeable with the missing named configuration; inspect their contents before reproducing the setup. [S005] [S006] [S011]

Control also has a nested SOEM dependency. Preserving the parent archive alone would omit both principal submodules and this further dependency, so a repository download is not by itself a complete reproducible environment. [S001] [S005] [S011]

Calibration is part of the workflow

The wiki calls for joint initialization at each power-on using alignment pieces, with motor offsets recorded by the controller. It separately describes comparing hanging hardware and simulated joint trajectories while adjusting the model. Those steps explain why a policy checkpoint alone is not enough to commission another build. No sensor calibration, motor tuning, trajectory test or walking trial was performed here. [S010]

Keep the license boundaries explicit

The parent repository has an MIT license. The separate pinned repositories have no license-named file in their inventories, while selected source headers include an MIT notice in the Teensy sketch and differing NVIDIA notices in simulation files. Do not infer that the parent's MIT license covers all submodule code, Isaac Gym, third-party meshes, manufacturing attachments or policy weights. Baseline/passive checkpoints and hardware ONNX policies are preserved as artifacts, but their specific reuse rights and execution remain unverified. [S003] [S011]

This is a source map for a research platform with useful hardware documentation. It is not a validated build pack, a fully resolved open-hardware license claim or a tested locomotion implementation. [S007] [S008] [S009] [S011]