Skip to content
Marvin Bot

Acknowledgments

Marvin Bot is designed from scratch, but it learns from many open projects, published specifications and tools. This file credits them. Nothing from them is copied into this repository unless a line below says so and the licence allows it.

Humanoid designs (ideas and reference figures)

Section titled “Humanoid designs (ideas and reference figures)”
  • Asimov 1: https://github.com/menloresearch/asimov-1 (hardware CERN-OHL-S-2.0, software GPL-2.0). The main reference for the overall layout: a 1.2 m, 35 kg, 25-DOF humanoid with 6-DOF legs, 5-DOF arms, waist yaw and a 2-DOF neck, a two-motor parallel-linkage ankle, and per-joint torque classes. Used for ideas and sizing only; no Asimov CAD, drawings or code are used here.
  • Roboto Origin: https://github.com/Roboparty/roboto_origin (GPL-3.0 at the top level; its hardware, description and training modules keep their own licences, listed in its LICENSES.md). An open 1.25 m, 34 kg, 23-DOF humanoid; reference for an all-machined structure, two-motor ankles with a parallel linkage, an off-the-shelf 48 V pack, a published BOM and per-joint actuator gains for Isaac Lab. Used for ideas and reference figures only.
  • RP1 (RoboParty): https://github.com/Roboparty/RP1_hardware (hardware CERN-OHL-W-2.0, documentation CC-BY-SA-4.0) and https://github.com/Roboparty/rp1_description. An open human-proportioned humanoid; its published assembly was studied for the human-form frame study of 2026-10-31 in docs/decisions.md (ankle motors with their axes fore and aft driving cranks and rods to the heel, hip roll behind the hip, hip yaw low in the thigh, a waist roll under the waist yaw, actuators left visible at the joints between covers). Ideas only; no RP1 CAD, drawings or files are used here.
  • Berkeley Humanoid Lite (UC Berkeley Hybrid Robotics): https://github.com/HybridRobotics/Berkeley-Humanoid-Lite and arXiv:2504.17249. Lessons on low-cost actuation, printed structures that route load through metal fasteners, and sim-to-real training.
  • ToddlerBot (Stanford University): https://github.com/hshi74/toddlerbot. Lessons on zero-point calibration and data collection for a low-cost open humanoid.
  • K-Scale Labs: open RobStride actuator drivers (https://github.com/kscalelabs/actuator, MIT) and an all-RobStride open humanoid (https://github.com/kscalelabs/kbot; hardware CERN-OHL-S-2.0, software MIT) were an inspiration only. No code or CAD from them is used: Marvin Bot’s RobStride driver is written from RobStride’s manuals.

Software and models

Section titled “Software and models”

  • NVIDIA Isaac Lab (https://github.com/isaac-sim/IsaacLab, BSD-3-Clause), Isaac Sim (NVIDIA licence, installed by the user, not redistributed), Newton and Isaac ROS: training environments and the onboard perception stack. The walking tasks’ structure and rewards, the PPO settings, the terrain generator config and DelayedDCMotor in sim/isaaclab/ are adapted from Isaac Lab’s G1 velocity tasks, rough-terrain config, quadruped foot-clearance term and DelayedPDActuator; those files keep Isaac Lab’s notice (LICENSES/BSD-3-Clause-IsaacLab.txt).

  • NVIDIA Isaac GR00T N1.7 (https://github.com/NVIDIA/Isaac-GR00T) and GR00T-WholeBodyControl (https://github.com/NVlabs/GR00T-WholeBodyControl): the planned vision-language-action model and the decoupled whole-body control approach.

  • MuJoCo (Google DeepMind, https://github.com/google-deepmind/mujoco, Apache-2.0): the simulation cross-checks and renders.

  • rsl_rl (ETH Zurich Robotic Systems Lab, https://github.com/leggedrobotics/rsl_rl, BSD-3-Clause): PPO training, used as a library.

  • ProtoMotions (NVIDIA, https://github.com/NVlabs/ProtoMotions, Apache-2.0): the motion pipeline in sim/motion/ runs its keypoint extraction, PyRoki retargeter (G1 version, with Marvin Bot’s link map and scaling), converters, motion library and motion-tracking training. Its keypoint convention (NVIDIA’s SOMA skeleton) is what convert_100style.py writes; its own sample motions (the BONES-SEED dataset, licensed only to non-profit academic institutions and small companies) are not used.

  • 100STYLE (I. Mason, S. Starke and T. Komura, “Real-Time Style Modelling of Human Locomotion via Feature-Wise Transformations and Local Motion Phases”, 2022; https://zenodo.org/records/8127870, CC-BY-4.0): the motion capture the first trackers learned from, still supported as an alternative source. Downloaded by the user, not redistributed; sim/motion/convert_100style.py reads it.

  • Kimodo (NVIDIA, https://github.com/nv-tlabs/kimodo, code Apache-2.0; the Kimodo-SOMA-RP model under the NVIDIA Open Model License, trained on Bones Studio’s motion capture): generates the human motions the trackers learn from (sim/motion/generate_kimodo.py), following its documented prompting advice and root-path constraints. Its text encoder is Meta Llama 3 8B Instruct (Meta Llama 3 Community License, gated) with LLM2Vec adapters (https://github.com/McGill-NLP/llm2vec, MIT). sim/motion/check_clips.py scores the clips with NVIDIA’s TMR-SOMA-RP-v1 text-motion retrieval model (NVIDIA Open Model License) and Kimodo’s benchmark metrics. Models are downloaded by each user under those licences; nothing from them is redistributed here.

  • PyRoki (https://github.com/chungmin99/pyroki, MIT): the kinematics optimiser behind the retargeter.

  • BeyondMimic (https://github.com/HybridRobotics/whole_body_tracking, arXiv:2508.08241): the motion-tracking formulation ProtoMotions’ tracker follows. GMR (https://github.com/YanjieZe/GMR): a possible alternative retargeter. The CMU Graphics Lab Motion Capture Database (http://mocap.cs.cmu.edu) was compared with 100STYLE and Kimodo as a motion source and is not used.

  • NVIDIA Cosmos (world models and open physical-AI datasets), GR00T-Mimic and SONIC (GR00T-WholeBodyControl): planned sources of synthetic manipulation data and of a general whole-body tracker.

  • Papers behind the planned skills in docs/roadmap.md: HoST (arXiv 2502.08378) and arXiv 2502.12152 (getting up), TWIST (arXiv 2505.02833) and HOMIE (arXiv 2502.13013) (teleoperation and loco-manipulation), Hiking in the Wild (arXiv 2601.07718) and PIM (arXiv 2411.14386) (stairs from depth), PhysHSI (arXiv 2510.11072) (carrying and sitting), PAiD (arXiv 2602.05310) and the RoboCup 2025 soccer work (kicking). Ideas only; no code from them is used.

  • mjlab (https://github.com/mujocolab/mjlab, Apache-2.0): its G1 velocity task showed the anti-sway terms used in the walking rewards (an upright term and tight hip and ankle roll while walking, no single-support reward).

  • Anny (NAVER, https://github.com/naver/anny, Apache-2.0): the parametric human body model that generated the body used as the shape reference in the human-form frame study (docs/decisions.md, 2026-10-31). The body was generated locally and is not redistributed here.

  • manifold3d (https://github.com/elalish/manifold, Apache-2.0) and trimesh (https://github.com/mikedh/trimesh, MIT): the code-CAD in cad/ (solids, booleans, meshes); fast-simplification (MIT) decimates the hand meshes. NumPy, PyYAML, Pillow, ONNX Runtime, matplotlib, imageio and pytest for the tools and checks; python-can (LGPL-3.0, optional, bench only) for the CAN bus; PyTorch, JAX and yourdfpy through the motion pipeline’s upstream tools.

  • Astro and Starlight (https://github.com/withastro/astro, https://github.com/withastro/starlight, MIT): the documentation website in site/. three.js (https://github.com/mrdoob/three.js, MIT): the 3-D model on the website and in the interactive guide, bundled from the pinned npm package at build time. The interactive guide in site/src/content/docs/guide/ cites every paper, datasheet, document and project it draws on, inline and on its sources page.

  • fTetWild (https://github.com/wildmeshing/fTetWild, MPL-2.0, through pytetwild, MPL-2.0, and PyVista, MIT), scikit-fem (https://github.com/kinnala/scikit-fem, BSD-3-Clause), PyPardiso (https://github.com/haasad/PyPardisoProject, BSD-3-Clause; Intel oneMKL PARDISO) and Rtree (MIT): the finite-element stress check in cad/fea.py (meshing, assembly, the sparse solver, point-in-part queries).

  • build123d (https://github.com/gumyr/build123d, Apache-2.0) on OCP (https://github.com/CadQuery/OCP, Apache-2.0) and Open CASCADE Technology (https://dev.opencascade.org, LGPL-2.1 with the OCCT exception): the machined parts’ B-rep models, their STEP files and the outlines they are cut from (cad/b3d.py).

  • bd_warehouse (https://github.com/gumyr/bd_warehouse, Apache-2.0): a planned source of parametric fasteners and bearings.

  • Claude and Claude Code (Anthropic): the project is developed with them, from research and design reviews to the code-CAD, simulation and training code and the documentation; commits they contributed to carry a Co-Authored-By: Claude line.

Component specifications

Section titled “Component specifications”
  • RobStride official manuals and specifications: https://github.com/RobStride/Product_Information. Actuator figures in description/actuators.yaml come from these manuals unless marked otherwise. The actuators’ STEP models from the same repository are downloaded by tools/fetch_vendor_cad.py (pinned to a commit); outlines and hole patterns in actuators.yaml are measured from them, and they are not redistributed here.
  • Feetech STS3250 datasheet (neck and wrist pitch servos) and SCS0009 figures (the hands’ finger servos).
  • AmazingHand by Pollen Robotics: https://github.com/pollen-robotics/AmazingHand (mechanical design CC-BY-4.0, code Apache-2.0). Marvin Bot’s hands are AmazingHand kits (bought from Seeed Studio, SKUs 100062181 and 100063642). The hand meshes in description/third_party/amazing_hand/ are derived from its CAD and MuJoCo models (merged, simplified, re-framed; see the README there) and stay under CC-BY-4.0; tools/build_hand_mesh.py rebuilds them. Its published mass and wrist-interface bolt pattern set the hand mount in cad/arm.py. No AmazingHand code is used.
  • Orbbec Gemini 335L product documentation (head depth camera), and its mesh and URDF from Orbbec’s ROS 2 driver, OrbbecSDK_ROS2 (orbbec_description, https://github.com/orbbec/OrbbecSDK_ROS2, Apache-2.0): tools/fetch_vendor_cad.py downloads them (pinned to a commit) and the camera’s outline, lens positions and tripod thread in cad/head.py are measured from them. They are not redistributed here.
  • NVIDIA Jetson AGX Orin developer kit (P3730) model: its outline and port positions place the computer in cad/electrical.py; the model is used locally and not redistributed.
  • Prices in bom/bom.xlsx (its group and Sourcing sheets) come from the vendors’ public listings, each dated in the workbook (no affiliation). The developer kit’s July 2026 price change is from CNX Software’s report: https://www.cnx-software.com/2026/07/22/nvidia-increases-the-price-of-jetson-modules-and-devkits-by-up-to-101/.
  • Human body proportions for the arm lengths (forearm 0.146 and hand 0.108 of height): D. A. Winter, Biomechanics and Motor Control of Human Movement, after Drillis and Contini (1966).

This page is generated from ACKNOWLEDGMENTS.md in the repository.