[far-future / robot concept] MEVIUS2 — open-source quadruped reference for home assistant locomotion stack #45

New issue

Open

opened 2026-06-02 20:39:46 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-06-02 20:39:46 -07:00

Owner

Summary

MEVIUS2 (https://github.com/haraduka/mevius2) is an open-source quadruped robot with a well-documented RL locomotion stack. MIT license. IEEE R&A paper. Relevant as a reference design for the FOSS/OSH home assistant robot concept (Labrador Systems equivalent).

Why it matters for CF

The CF home assistant robot concept targets wheeled-base + arm form factor, not quadruped, but MEVIUS2's software stack is directly applicable:

Locomotion: legged_gym RL policy training, exported to PyTorch, runs on-device via ROS2
Simulation: MuJoCo + IsaacGym for sim-to-real transfer
Sensing: FAST_LIO SLAM, LiDAR (Livox), IMU, elevation_mapping_cupy for terrain awareness
Actuation: RobStride03 motors via CAN bus — affordable and available

Design note: camera placement

MEVIUS2 (and most quadrupeds) mount cameras at chest/nose level. For a CF home assistant, camera should be high-center (top of the body volume, forward-facing) for:

Unoccluded ground plane view — no limb/wheel interference
2-3 stride lookahead for terrain anticipation
Stable egocentric perspective that RL locomotion policies can train on cleanly
Matches egocentric video training data distributions (cf. SynData dataset)

Hardware notes

Sheet metal chassis: manufacturable at low cost, OSH-friendly
RobStride03: 12 actuators for quadruped; wheeled base would need fewer
CAD files on Google Drive (linked from repo)

Status

Far-future backlog concept. No active development. File as watch/reference alongside the Labrador-equivalent home assistant concept.

License: MIT
Stars: 114

## Summary MEVIUS2 (https://github.com/haraduka/mevius2) is an open-source quadruped robot with a well-documented RL locomotion stack. MIT license. IEEE R&A paper. Relevant as a reference design for the FOSS/OSH home assistant robot concept (Labrador Systems equivalent). ## Why it matters for CF The CF home assistant robot concept targets wheeled-base + arm form factor, not quadruped, but MEVIUS2's software stack is directly applicable: - **Locomotion:** legged_gym RL policy training, exported to PyTorch, runs on-device via ROS2 - **Simulation:** MuJoCo + IsaacGym for sim-to-real transfer - **Sensing:** FAST_LIO SLAM, LiDAR (Livox), IMU, elevation_mapping_cupy for terrain awareness - **Actuation:** RobStride03 motors via CAN bus — affordable and available ## Design note: camera placement MEVIUS2 (and most quadrupeds) mount cameras at chest/nose level. For a CF home assistant, camera should be **high-center** (top of the body volume, forward-facing) for: - Unoccluded ground plane view — no limb/wheel interference - 2-3 stride lookahead for terrain anticipation - Stable egocentric perspective that RL locomotion policies can train on cleanly - Matches egocentric video training data distributions (cf. SynData dataset) ## Hardware notes - Sheet metal chassis: manufacturable at low cost, OSH-friendly - RobStride03: 12 actuators for quadruped; wheeled base would need fewer - CAD files on Google Drive (linked from repo) ## Status Far-future backlog concept. No active development. File as watch/reference alongside the Labrador-equivalent home assistant concept. - License: MIT - Stars: 114