RMiner — Robot Manipulation Benchmark Datasets

Teleoperation datasets collected for RMiner, a benchmark of mining / construction manipulation tasks performed with a Unitree G1 (29-DoF + Dex1 grippers) in NVIDIA Isaac Sim 5.1. All datasets are stored in LeRobot format (v3.0 unless noted), recorded at 30 FPS with head, left-wrist and right-wrist camera streams plus proprioceptive state and action.

Code, environment configuration and task definitions: https://github.com/Yy12136/RMiner

Layout

datasets/<task_directory>/<lerobot_version>/<dataset_name>/

Each dataset is a self-contained LeRobot dataset (meta/, data/, videos/).

Task	Dataset	Version	Episodes	Frames	Notes
Task01 Tabletop Grasp	g1_tabletop_grasp_task/lerobot_v30/detonator_wire_insertion	v3.0	250	221,315	main / full
Task01 Tabletop Grasp	.../detonator_wire_insertion_50_20260822	v3.0	51	44,983	50-episode subset
Task01 Tabletop Grasp	.../detonator_wire_insertion_250_video_only_20260823	v3.0	250	221,315	video-only variant
Task01 Tabletop Grasp	.../detonator_wire_insertion_replay_ready	v3.0	50	54,464	replay-augmented
Task01 Tabletop Grasp	g1_tabletop_grasp_task/lerobot_v21/detonator_wire_insertion	v2.1	1	1,540	legacy sample
Task02 Roof Bolt Sleeve	g1_roof_bolt_sleeve_task/lerobot_v30/roof_bolt_sleeve_installation	v3.0	51	48,368	base
Task02 Roof Bolt Sleeve	.../roof_bolt_sleeve_installation_final	v3.0	250	238,015	main / full
Task02 Roof Bolt Sleeve	.../roof_bolt_sleeve_installation_replay_ready	v3.0	50	48,498	replay-augmented
Task03 Bin Pick & Place	g1_bin_pick_place_task/lerobot_v30/sam3d_metal_component_to_randomized_bin	v3.0	50	45,775	base
Task03 Bin Pick & Place	.../sam3d_metal_component_to_randomized_bin_final	v3.0	250	228,875	main / full
Task03 Bin Pick & Place	.../sam3d_metal_component_to_randomized_bin_replay_ready	v3.0	50	45,775	replay-augmented
Task04 Hand Cart Push	g1_hand_cart_push_task/lerobot_v30/hand_cart_push	v3.0	50	14,750	base
Task04 Hand Cart Push	.../hand_cart_push_final	v3.0	250	73,750	main / full
Task04 Hand Cart Push	.../hand_cart_push_replay_ready	v3.0	50	17,364	replay-augmented
Task05 Valve Handwheel	g1_valve_handwheel_task/lerobot_v30/valve_handwheel_rotation	v3.0	51	37,116	base
Task05 Valve Handwheel	.../valve_handwheel_rotation_final	v3.0	250	182,300	main / full
Task05 Valve Handwheel	.../valve_handwheel_rotation_replay_ready	v3.0	50	36,498	replay-augmented
Task06 Wall-hole Detonator	g1_wall_hole_detonator_task/lerobot_v30/wall_hole_loaded_detonator_insertion	v3.0	50	31,638	base
Task06 Wall-hole Detonator	.../wall_hole_loaded_detonator_insertion_final	v3.0	250	158,190	main / full
Task06 Wall-hole Detonator	.../wall_hole_loaded_detonator_insertion_replay_ready	v3.0	50	31,638	replay-augmented
Hand-cart Rolling Demo	g1_hand_cart_rolling_demo_task/lerobot_v30/hand_cart_rolling_demo	v3.0	—	—	incomplete: raw .staging/ videos only, no meta/info.json

3D assets

The same repository also hosts the USD/USDZ/texture assets required to run the benchmark, under assets/assemble/:

Path	Contents
assets/assemble/assets-new/	shared scenes, robots, objects (incl. RealScene04)
assets/assemble/objects/	task-specific USD objects (hand cart, valve, drill, ...)
assets/assemble/task/	per-task assets for w1_mobile_drilling_task and quest_dex3_hand_test
assets/assemble/third_party/unitree_sim_isaaclab/assets/	robot assets used by the simulator

Download only the assets:

hf download lyang116/RMiner --repo-type dataset \
  --include "assets/**" --local-dir ./RMiner-assets

The code repository's scripts/download_assets.sh does this and installs the files into the expected assemble/ layout automatically.

Download

# whole repository (recommended: git-xet)
git xet install
git clone https://huggingface.co/datasets/lyang116/RMiner

# or a single task with the Python client
hf download lyang116/RMiner \
  --repo-type dataset \
  --include "datasets/g1_valve_handwheel_task/**" \
  --local-dir ./RMiner

If you already use the RMiner repository layout, copy a dataset back into place:

cp -a datasets/g1_valve_handwheel_task/lerobot_v30/valve_handwheel_rotation \
      RMiner/assemble/task/g1_valve_handwheel_task/data/lerobot_v30/

Dataset fields

LeRobot v3.0 default schema for all tasks:

• observation.state — G1 joint positions (29 DoF) + gripper state
• action — target joint positions / gripper commands
• observation.images.head, observation.images.left_wrist, observation.images.right_wrist
• task — natural-language instruction
• timestamp, frame_index, episode_index, index

See each dataset's meta/info.json, meta/episodes.jsonl and meta/tasks.jsonl for the exact feature list and episode metadata.

License

Code and datasets are released under Apache-2.0. Third-party robot/scene assets retain their original licenses (see the code repository for details).