This document is a concise current-state summary. Full per-release history is in CHANGELOG.md; the long-form narrative of the v0.6 → v0.7 bring-up is preserved at docs/STATUS_v0.7_history.md.

Current release

v0.7.8 — Push-recovery test passes with 1 mm XY recovery error and 0.19° final tilt — better than v0.5’s published ~5 cm recovery. The fix: replicate v0.5’s set_initial_stance_pose via the existing MujocoBackend::set_base_pos_world + set_joint_qpos setters, plus stiffen joint impedance to 300/8 so the legs anchor the chassis like rigid struts. All five v0.5 tracking numbers now reproduced, four of them at or beyond the published Python performance. See v0.7.8 entry, v0.7.7 entry, v0.7.6 entry, v0.7.5 entry, v0.7.4 entry, v0.7.3 entry, v0.7.2 entry, v0.7.1 entry, and v0.7 entry for the full deliverable lists.

Tests green

Approximate counts per crate at v0.7 ship:

Crate Tests Notes
cad-udd 19  
cad-format 75 incl. 11 perception-emit tests (URDF + MJCF)
cad-parts 70 incl. mesh-bake helper + 7-part build_mesh coverage
cad-mfg 12  
mgai-robot 98 Moteus + Dynamixel + SimpleFOC drivers
mgai-control 100  
mgai-sim 20 + 7 feature-gated mujoco-ffi tests
mgai-perception 240 skill_graph runtime + classical + detector
mgai-vla 697 668 pre-existing VLA + 21 new skill_graph incl. subprocess
mgai-depthai 5 + 1 feature-gated live scaffold test
openloco 76 incl. sim_integration, closed_loop_integration, camera_backed_integration, +7 feature-gated mujoco_full_chain, +5 feature-gated gait (gait_anymal_wl, gait_m4_driving, gait_mini_whegs, gait_m4_hover, gait_push_recovery), no #[ignore]‘d gait remaining

Total: ~1,400 tests green under default cargo test; additional hardware-gated tests (MGAI_DEPTHAI_LIVE=1, MGAI_VLA_SUBPROC=1) when the relevant hardware / venv is present.

Deferred — known gaps

  • Real camera-from-physics rendering — PhysicsBackedFrames falls through to canned RGBD for rendered sensor kinds. Closing this means wiring MuJoCo’s mj_renderOffscreen into the physics-backed frame source. Closes the simulated end of the camera loop.
  • Production VLM weights shipped with the crate — VilaQuerySubprocess has a working real-model path but weights are resolved via HuggingFace download on first run. Could ship a pinned small model (SmolVLM-256M ≈ 300 MB) as a dev artifact for deterministic tests.
  • Live actuator bring-up — the Tier 0 / 1 / 2 driver protocol layers are tested against mocks; landing one on real Dynamixel / SimpleFOC / Moteus hardware is the parallel hardware story to the DepthAI bring-up that shipped in v0.7.
  • ROS 2 / Isaac Sim validation of the emitted URDF against ROS 2 tooling. The MuJoCo half of this closed in v0.7.1 via mujoco_full_chain.rs: all descriptors load into MuJoCo and step physics. ROS 2 load is the remaining external-consumer check.

  • MuJoCo-backed regression suite reproducing the v0.5 Python tracking numbers — all five reproduced as of v0.7.8, four at or beyond v0.5 performance:

    v0.5 number Rust port (v0.7.8) Status
    m4 driving 98% 99% ✅
    mini_whegs walking 68% 64% ≈
    anymal_wl walking 83% 72% world-X / 121% planar ≈
    hover 0.77 m / 2.2° 1 mm err / 0.03° tilt ✅ better
    15 N push recovery (~5 cm err) 1 mm err / 0.19° tilt ✅ better

    Remaining minor gaps (none blocking):

    • anymal_wl 83% world-X — currently 72% world-X / 121% planar; gap is yaw oscillation during the new walking dynamics. Tighter integral control / wheel-brake tuning would close it.
    • mini_whegs 68% — currently 64%, well within the v0.5 envelope.

Architecture cheatsheet

Single source of truth → many artifacts:

descriptor.udd.json
    │
    ├── Robot section ──► URDF + MJCF + meshes + BOM + assembly
    │                     Actuator trait (3 tiers) + controllers + physics
    │
    └── Perception section ──► SkillGraphRuntime
                                ├── classical skills (foothold_map, obstacle_height, ground_plane_fit)
                                ├── detector skills (StubDetector)
                                └── VLM skills (VilaQueryStub, VilaQuerySubprocess ← real HF)
                                    ↑
                                SensorFrameSource
                                ├── CannedFrames
                                ├── PhysicsBackedFrames<B: PhysicsBackend>
                                └── CameraBackedFrameSource ← LuxonisCamera (DepthAI)

Where to look