skyfull-chamonix-lookat-v1

Pointing / look-at dataset for the SO-101 robotic arm over virtual Chamonix terrain.

Generated with Skyfull — each episode: a random avalanche / victim event fires on the terrain, the virtual arm rotates to point its sensor at the event location, joint angles are recorded.

What this is for

Sim-to-real transfer: train the arm to point at any location on a 2D/3D map representation. Transfer to the real SO-101 with a single coordinate-frame calibration (3 reference points → affine transform).

Data schema

Each row (9 float32 values):

Column	Unit	Description
lon	°	Event longitude
lat	°	Event latitude
alt_m	m	Event altitude (terrain surface)
x_m	m	Event x in arm frame (MuJoCo metres)
y_m	m	Event y in arm frame
z_m	m	Event z in arm frame
theta0	rad	SO-101 joint 0 — pan (Z-axis rotation)
theta1	rad	SO-101 joint 1 — shoulder tilt (-Y-axis)
theta2	rad	SO-101 joint 2 — elbow (0 = fully extended)

Tile

Region	Chamonix, Haute-Savoie, France
Bounds	lat [45.86, 45.975] lon [6.8, 6.99]
Episodes	5,000
Simulator	MuJoCo + SO-101 3-DOF arm (L0=65mm L1=105mm L2=105mm L3=55mm)

Quick start

import numpy as np
from huggingface_hub import hf_hub_download

ep = np.load(hf_hub_download("Ethgar/skyfull-chamonix-lookat-v1", "episodes.npy", repo_type="dataset"))
print(ep.shape)        # (5000, 9)
print(ep[0])           # [lon, lat, alt_m, x_m, y_m, z_m, θ0, θ1, θ2]

Companion model

Ethgar/skyfull-lookat-chamonix-v1 — 4.5k-param LookAtNet trained on this data. Mean angular error < 0.2°.