dim-go2-sim

A simulated Unitree Go2 for dimOS, as one module: cmd_vel in; a rolling Livox Mid-360, a textured front camera, IMU, odometry, tf and contacts out. It drops into a blueprint where a real Go2 would be.

pip install dim-go2-sim

from dimos.core.coordination.blueprints import autoconnect
from dim_go2_sim import Go2Sim

blueprint = autoconnect(
    Go2Sim.blueprint(scene={"package": "supermarket", "start": (18.5, 10.0, -0.17)}),
    MyNavigation.blueprint(),  # anything that reads lidar/odom and writes cmd_vel
)

With no scene it stands on bare ground. A scene is a cooked dimOS scene package by name (scene.package), an InteriorGS home (scene.package="interiorgs/839920"), a real-world splat scan (scene.package="supersplat/3f89bbd3", or scene.scan=<a .ply>), or an MJCF of the collision geometry (scene.mjcf) and optionally what the camera draws (scene.visual): a textured glTF, or Gaussian splats (.nurec, a 3DGS .ply).

InteriorGS homes

Go2Sim.blueprint(scene={"package": "interiorgs/839920"})  # any of the 1000 scene ids

A furnished home from SpatialVerse's InteriorGS, drawn by the camera from its Gaussian splats. The first run downloads its splats and collision meshes (100-350 MB, from SAGE-3D's USDZ and collision mesh mirrors) and cooks them into physics.cache_dir in under a minute. The robot starts at the most open spot on the floor, facing the longest clear view; the cooked scene keeps up to 8 such spawn poses (dim_go2_sim.interiorgs.load_scene(id).spawns).

Splat scans

Go2Sim.blueprint(scene={"package": "supersplat/3f89bbd3"})  # a SuperSplat scene by id
Go2Sim.blueprint(scene={"scan": "my_home.ply", "scan_scale": 1.0})  # a 3DGS PLY or SOG meta.json

A Gaussian-splat scan of a real home: the camera draws its splats, and the robot and lidar collide with geometry built from them. The default scan is an NJ ranch home (2.7M splats, 31 MB, downloaded from SuperSplat on first use). A scan has no collision mesh and an arbitrary frame and scale, so the first run cooks it (a few seconds):

• Level. The floor, the densest layer at the bottom of the scan, is fitted by RANSAC and turned to z = 0; the walls (thin upright splats) are turned square to x and y.
• Metres. A scan's units are arbitrary: scene.scan_scale gives metres per unit. A known SuperSplat scene has a measured one (dim_go2_sim.supersplat.SCALES); the ranch home's is 0.92, from its US-standard fittings: its kitchen counters are 0.985 units high and stand 36 in (0.914 m), its interior door openings 2.31 units and an 80 in door's ~2.05 m (0.928 and 0.89), and its walls, cut off at the ceiling, then reach 2.3 m, under an 8 ft ceiling.
• Collisions. Opaque splats under 15 cm across are voxelised at 5 cm (two to a voxel), specks under 30 voxels dropped as floaters, and a column solid for 0.6 m of the metre below 2 m filled down to the floor (walls' feet are thin in a scan). As with InteriorGS, the voxels up to 2 m are merged into boxes for MuJoCo over a floor plane, and the lidar casts against the voxels' faces and the floor under the scan. Spawns are picked on open floor.

Only a scan's degree-0 (view-independent) colour is kept: higher spherical harmonics would have to be turned with it. Splats from a 3DGS PLY are drawn as the InteriorGS ones are, evaluated in 3D.

Streams

The URDF

As dimOS's UrdfManager shares one: each mesh is named package://dim_go2_sim/<path>?hash=<sha256 of the file>, and that name on mesh_request gets the file on mesh_info (named with its current hash) unless the request ends &have=<that hash>. The root link is world; base_link hangs off it by a floating joint and each link is posed by its tf frame. The scene is fixed at world: scene_mesh is its collision surfaces, simplified to urdf.scene_faces triangles in STL tiles (scenes/<scene>/mesh_000.stl, ...), and for a splat scene (an InteriorGS home, a scan) scene_splats is its Gaussian splats as one SPZ file (scenes/interiorgs_<id>/splats.spz, world frame, Z up). The files are built once per scene into physics.cache_dir. urdf.enabled=false turns all of it, and the leg tf, off.

Config

Go2SimConfig has a section per part, each overridable from a blueprint or the dimOS CLI:

• scene: package, scan, scan_scale, mjcf, visual, visual_y_up, start (x, y, yaw; default the scene's first spawn), floor_z, floor_top_z
• physics: timestep, fitted, actuator_tau, policy, track_velocity, command_timeout, max_velocity, realtime, cache_dir
• lidar: enabled, rate_hz, xyz, pitch_deg, point_rate_hz, downsample, min_range, max_range, noise, dropout, self_occlusion, motion_bins_hz, seed
• camera: enabled, rate_hz, width, height, fovy_deg, xyz, pitch_deg, near, far, detail_distance, simplify, max_texture_size, ambient, sun, background
• urdf: enabled, period_s, scene_faces, tile_faces, splats_sh_degree
• odom_hz, imu_hz, frame_prefix

See src/dim_go2_sim/config.py for what each means and its default.

How it works

• Walking. MuJoCo steps the Go2 under the open rl_sar robot_lab policy at 50 Hz, with joint damping, friction, armature, masses, foot friction and a 15 ms motor lag fitted to real Go2 runs under that policy. The policy alone ignores slow commands and drifts when told to stand, so an integral loop on the true body velocity trims its command, as the Go2's sport mode tracks it (physics.track_velocity).
• Lidar. The Mid-360's four-channel Fourier firing pattern, 200k points a second, cast in 5 ms slices from wherever the head is then, with range noise and grazing-angle dropout. The static scene is cast through Embree; the robot's own body, through MuJoCo. A cooked store of 48k geoms scans at the full 10 Hz.
• Camera. pygfx draws the scene's textured glTF in a child process, its meshes merged by material and 4 m cell (a 47k-node store: 0.6 → 60+ fps on an M-series Mac, every object drawn). It draws the scene only, not the robot. On a software GPU (Mesa's lavapipe, e.g. in Docker: install mesa-vulkan-drivers) it simplifies the scene and draws small objects only within 4 m, to keep a store at several frames a second (camera.simplify, camera.detail_distance). A store's full-size textures overflow an 8 GB GPU; camera.max_texture_size (e.g. 2048) halves larger ones until they fit.
• Splat camera. Gaussian splats are drawn with wgpu (Metal, Vulkan or lavapipe), sorted far to near each frame, each pixel's ray evaluated against the splat in 3D as 3DGUT (which made the InteriorGS splats) renders it, with degree-3 spherical harmonics. An InteriorGS home (1.2M splats) at 640x480: ~25 fps on an M-series Mac, ~45 on an RTX 5070 laptop, ~2 on lavapipe.
• InteriorGS collisions. MuJoCo collides only with convex shapes and the homes' meshes are open surfaces, so they are voxelised at 5 cm up to 2 m above the floor and the voxels merged into boxes (a few thousand per home); the floor is a plane. The lidar casts against the meshes themselves.
• The composed world is compiled once and cached (physics.cache_dir, default ~/.cache/dim_go2_sim); a store takes about 2 minutes the first time.

Development

uv venv --python 3.12 && uv pip install -e . pytest
pytest

The tests on a real InteriorGS home and the ranch scan run once they are cooked, or with DIM_GO2_SIM_DOWNLOAD=1.

Licenses

Apache-2.0. Bundled data keeps its own terms: the Go2 model (Unitree, BSD-3, data/go2/LICENSE), the walking policy (rl_sar, Apache-2.0, data/policy/NOTICE), and the Mid-360 pattern (data/mid360/NOTICE.md).

InteriorGS scenes are not bundled: they are downloaded on first use, and keep their own license, CC BY-NC 4.0 (non-commercial use only). InteriorGS is by SpatialVerse; the USDZ and collision-mesh conversions are SAGE-3D's (paper, code).

The ranch home scan is not bundled either: it is downloaded from SuperSplat on first use. "Home Scan (Creation process in description)" is by Isaiah Sweeney (luxury_scans), licensed CC BY 4.0. The sim levels, scales and voxelises it, and keeps only its degree-0 colour.