Documentation index and glossary
| Document | What it is |
|---|---|
| plan.md | Phases, their pass checks and the repository’s planned layout |
| roadmap.md | Where each phase stands, open items, skills planned beyond walking |
| requirements.md | What the robot must and should do |
| decisions.md | Every design decision, newest first, with reasons and alternatives (older entries describe the design as it was then) |
| safety.md | Safety rules for benches, batteries and walking tests |
| design/concept.md | The v0.2 concept (historical) |
| design/stairs.md | Stair climbing: status, open issues and options |
| design/assembly_wiring.md | Modules, assembly order, power and wiring |
| ../sim/isaaclab/README.md | Training the walking and stair policies, and every run’s results |
| ../sim/motion/README.md | The motion-tracking pipeline |
| ../bench/README.md | Actuator bench tests (phase 2) |
| ../bom/README.md | Bill of materials and how its prices are sourced |
Glossary
Section titled “Glossary”- DOF: degree of freedom, one powered joint axis. Marvin Bot has 27, plus 8 in each hand.
- QDD (quasi-direct drive): an actuator with a low gear ratio (about 9:1 here), so it can be back-driven and sense force through its current.
- RobStride RS00/RS10P/RS03/RS04: the actuator models used, from small (wrist) to large (hips, knees).
- URDF, MJCF, USD: robot description formats: URDF is generated here from
description/; MuJoCo reads URDF or MJCF; Isaac Sim uses USD. - Code-CAD: parts written as Python code (
cad/) instead of drawn in a CAD program, so they regenerate from the description. - SF (safety factor): yield strength divided by the calculated stress; the strength check targets 2.
- Rated vs peak torque: a motor can hold its rated torque continuously; peak torque only briefly. RMS torque over a motion is compared with the rated value for heat.
- CoM: centre of mass.
- IMU: inertial measurement unit (gyro and accelerometer) in the pelvis.
- BMS: battery management system.
- CAN: the bus the actuators are commanded over.
- PPO: the reinforcement-learning algorithm used to train the walking policies.
- Policy: the trained neural network that turns sensor readings into joint targets.
- Flat v7, rough v9: the 7th flat-ground and 9th rough-terrain training runs; each resumed the one before after a change.
- Sim-to-real: making a policy trained in simulation work on the real robot.
- Cross-check: replaying a policy trained in Isaac Lab in a second simulator (MuJoCo) to catch simulator-specific tricks.
- Motion tracking (BeyondMimic): training a policy to follow a reference motion, such as a retargeted human walk.
- Retargeting: converting a human motion to the robot’s body and joints.
- VLA: vision-language-action model (GR00T N1.7 here), which turns camera images and an instruction into arm motions.
This page is generated from docs/README.md in the repository.