Robot videos: links

Links for everything in the robot videos: what you need to try it yourself and where the numbers come from. Newest video first. The videos are in Turkish; on Instagram you can DM me the keyword and I'll send the same links.

NVIDIA SimReady: why are 2,677 objects free? · DM: SIM

SimReady assets are 3D objects ready for robot simulation: besides the looks, the file carries mass, friction, a collision shape and labels. NVIDIA opened them on Hugging Face, free and ungated.

• Search and download in the browser: SimReady Central (HF Space) · try "coffee mug", "forklift", "franka".
• Dataset: nvidia/simready-assets (mostly CC BY 4.0, licence per asset)
• Announcement: NVIDIA on the Hugging Face blog (21 Oct 2026)
• What SimReady is: NVIDIA glossary · simready-foundation (Apache-2.0)
• Why simulation: NVIDIA on synthetic data

How I brought one into MuJoCo (4 steps)

MuJoCo from PyPI does not open USD directly, so I wrote the shape to OBJ with usd-core and gave that to MuJoCo.

1. Find the asset in the Space and press Download (the mug is ~15 MB).
2. pip install usd-core numpy
3. Select the package's MuJoCo variant and write the shape to OBJ:

from pxr import Usd, UsdGeom

stage = Usd.Stage.Open("simready_usd/sm_dishware_mug_coffee_a01_01.usd")
root = stage.GetDefaultPrim()
vs = root.GetVariantSets()
if vs.HasVariantSet("MuJoCo"):
    vs.GetVariantSet("MuJoCo").SetVariantSelection("Enabled")
cache = UsdGeom.XformCache()
lines, base = [], 1
for prim in Usd.PrimRange(root, Usd.TraverseInstanceProxies(Usd.PrimAllPrimsPredicate)):
    if not prim.IsA(UsdGeom.Mesh):
        continue
    mesh = UsdGeom.Mesh(prim)
    m = cache.GetLocalToWorldTransform(prim)
    pts = [m.Transform(p) for p in mesh.GetPointsAttr().Get()]
    lines += [f"v {p[0]} {p[1]} {p[2]}" for p in pts]
    counts = mesh.GetFaceVertexCountsAttr().Get()
    idx = mesh.GetFaceVertexIndicesAttr().Get()
    k = 0
    for c in counts:
        lines.append("f " + " ".join(str(base + i) for i in idx[k:k + c]))
        k += c
    base += len(pts)
open("mug.obj", "w").write("\n".join(lines))

4. Reference it in MJCF as <mesh file="mug.obj"/> and set mass and friction from the USD's physics:mass and friction values. Units are metres, z up, same as MuJoCo.

PaceVLA: same model, done 1.8× sooner · DM: PACE

• Paper (not peer-reviewed, no code): arXiv 2610.25961 · project page
• Author's announcement: X, @rakybond007 (23 Oct 2026)
• Related idea, training-free: PACE, arXiv 2606.00537
• The arm in the video is our own MuJoCo animation (Menagerie franka_emika_panda, Apache-2.0), not PaceVLA itself.

The duck and the half-centimetre step · DM: ORDEK

• Walk the duck in simulation: Open_Duck_Playground · Open_Duck_Mini
• Training on a Mac (our fork): fport/open-duck-mac
• Microduck (camera, football): Pollen Robotics announcement · open policies (Apache-2.0)
• NVIDIA's parkour claim: NVIDIA developer blog

Reachy Mini · DM: REACHY

• Try it without buying (official simulation): getting started
• SDK (Apache-2.0): pollen-robotics/reachy_mini
• Announcement: Hugging Face blog · A critical review: Jeff Geerling

Unitree G1 and Go2: training in simulation · DM: G1

• Training environments: MuJoCo Playground · unitree_rl_gym
• Robot models: MuJoCo Menagerie

A chat model on a robot arm · DM: KOL

• Evaluation: robocurve/inspect-robots · all runs · paper