office-rover-2task-raw

LeRobot v2.1 dataset for a single SO-101 arm: "pick up the red / green cube and put it in the box" with both cubes on the table. Two language-conditioned tasks (instructions in Russian). Built for fine-tuning NVIDIA Isaac GR00T N1.7 on one 24 GB RTX 4090 — recipe, patches and 1800 evaluated attempts: https://github.com/VShirokun/gr00t-on-4090

What is in it: the raw, un-engineered demonstrations: cubes in fixed orientation, 240×320 cameras. Trained as-is, GR00T N1.7 reached 386/600 = 64.33 % (CI 60.4–68.2) where a 450M VLA scored 0–6 %.

episodes	500
frames	69000 at 30 fps
cameras	front, wrist — 240×320
state / action	6 joints (5 + gripper), absolute targets
tasks	2

How it was made — read before you trust it. Simulation only: MuJoCo with the official SO-101 model from mujoco_menagerie. Demonstrations come from a scripted operator (inverse kinematics), not a human teleoperator. Cube positions are random over a 12×30 cm zone; success is judged by physics (named cube in the box, other cube untouched). The policy never sees cube coordinates. Real lighting, glare, friction and servo backlash are not represented — the sim-to-real gap is unmeasured.

Loads directly: LeRobotDataset("VShirokun/office-rover-2task-raw"). Converted from LeRobot v3.0 with the upstream scripts/lerobot_conversion/convert_v3_to_v2.py.

License: CC BY 4.0. Please cite the repository above.