LIBERO

LIBERO is a benchmark for **knowledge transfer in lifelong robot learning**, and its defining feature is experimental control rather than scale. Three of its suites hold everything fixed except one factor, so a policy's failure can be attributed to a specific kind of distribution shift instead of to "the scene changed". **The five suites.** - **LIBERO-Spatial** (10 tasks) — identical objects, different spatial relationships. The instruction is what disambiguates them: *pick up the black bowl next to the cookie box* vs *…on the ramekin*. - **LIBERO-Object** (10 tasks) — identical layout, a different object each time. - **LIBERO-Goal** (10 tasks) — identical objects and layout, a different goal each time. - **LIBERO-Long / LIBERO-10** (10 tasks) — long-horizon tasks that entangle all three factors, e.g. *turn on the stove and put the moka pot on it*. - **LIBERO-90** (90 tasks) — a broad pretraining pool drawn from kitchen, living-room and study scenes. LIBERO-90 and LIBERO-Long together are referred to as LIBERO-100. **What is actually in the release.** Exactly 50 human-teleoperated demonstrations per task — verified file by file, not assumed — giving **6,500 demonstrations and 1,007,618 transitions** across 130 HDF5 files totalling 93.5 GiB. At the 20 Hz control rate that is about 14 hours of interaction. Tasks are defined in PDDL-style BDDL files, so the suite is designed to be extended rather than frozen. **Observations and actions.** Each timestep carries an `agentview_rgb` and an `eye_in_hand_rgb` frame at 128×128, plus `joint_states` (7), `gripper_states` (2), `ee_pos` (3), `ee_ori` (3) and `ee_states` (6). Actions are 7-D OSC_POSE deltas. Full MuJoCo simulator state (`states`, 92-D) and the per-demo `init_state` are stored alongside, so any episode can be replayed deterministically — which is what every downstream re-render depends on. **Which copy to use.** This is worth knowing before downloading 93 GiB. The canonical HDF5 release is the one described above. Most vision-language-action work does **not** train on it directly, but on OpenVLA's regenerated `modified_libero_rlds`, which replays the demos in simulation and drops no-op-only and unsuccessful episodes — retaining **1,693 of 2,000 episodes (84.7%)** across Spatial, Object, Goal and Long. NVIDIA's LeRobot v3 conversion applies the same filtering and additionally covers LIBERO-90, keeping **5,614 of 6,500 episodes (86.4%)** and 842,714 of 1,007,618 frames. Physical Intelligence's LeRobot v2 copy is the four-suite filtered set (1,693 episodes / 273,465 frames). Numbers reported against "LIBERO" are usually against a filtered subset, not this one. **Licensing.** The authors' README licenses the codebase under MIT and the datasets under **CC BY 4.0**; the Hugging Face data mirror is labelled Apache-2.0. The CC BY 4.0 statement is the one that speaks specifically to the data and is used here.

language · LIBERO is a benchmark for **knowledge transfer in lifelong robot learning**, and its defining feature is experimental control rather than scale. Three of its suites hold everything fixed except one factor, so a policy's failure can be attributed to a specific kind of distribution shift instead of to "the scene changed". **The five suites.** - **LIBERO-Spatial** (10 tasks) — identical objects, different spatial relationships. The instruction is what disambiguates them: *pick up the black bowl next to the cookie box* vs *…on the ramekin*. - **LIBERO-Object** (10 tasks) — identical layout, a different object each time. - **LIBERO-Goal** (10 tasks) — identical objects and layout, a different goal each time. - **LIBERO-Long / LIBERO-10** (10 tasks) — long-horizon tasks that entangle all three factors, e.g. *turn on the stove and put the moka pot on it*. - **LIBERO-90** (90 tasks) — a broad pretraining pool drawn from kitchen, living-room and study scenes. LIBERO-90 and LIBERO-Long together are referred to as LIBERO-100. **What is actually in the release.** Exactly 50 human-teleoperated demonstrations per task — verified file by file, not assumed — giving **6,500 demonstrations and 1,007,618 transitions** across 130 HDF5 files totalling 93.5 GiB. At the 20 Hz control rate that is about 14 hours of interaction. Tasks are defined in PDDL-style BDDL files, so the suite is designed to be extended rather than frozen. **Observations and actions.** Each timestep carries an `agentview_rgb` and an `eye_in_hand_rgb` frame at 128×128, plus `joint_states` (7), `gripper_states` (2), `ee_pos` (3), `ee_ori` (3) and `ee_states` (6). Actions are 7-D OSC_POSE deltas. Full MuJoCo simulator state (`states`, 92-D) and the per-demo `init_state` are stored alongside, so any episode can be replayed deterministically — which is what every downstream re-render depends on. **Which copy to use.** This is worth knowing before downloading 93 GiB. The canonical HDF5 release is the one described above. Most vision-language-action work does **not** train on it directly, but on OpenVLA's regenerated `modified_libero_rlds`, which replays the demos in simulation and drops no-op-only and unsuccessful episodes — retaining **1,693 of 2,000 episodes (84.7%)** across Spatial, Object, Goal and Long. NVIDIA's LeRobot v3 conversion applies the same filtering and additionally covers LIBERO-90, keeping **5,614 of 6,500 episodes (86.4%)** and 842,714 of 1,007,618 frames. Physical Intelligence's LeRobot v2 copy is the four-suite filtered set (1,693 episodes / 273,465 frames). Numbers reported against "LIBERO" are usually against a filtered subset, not this one. **Licensing.** The authors' README licenses the codebase under MIT and the datasets under **CC BY 4.0**; the Hugging Face data mirror is labelled Apache-2.0. The CC BY 4.0 statement is the one that speaks specifically to the data and is used here.

LIBERO is a benchmark for **knowledge transfer in lifelong robot learning**, and its defining feature is experimental control rather than scale. Three of its suites hold everything fixed except one factor, so a policy's failure can be attributed to a specific kind of distribution shift instead of to "the scene changed". **The five suites.** - **LIBERO-Spatial** (10 tasks) — identical objects, different spatial relationships. The instruction is what disambiguates them: *pick up the black bowl next to the cookie box* vs *…on the ramekin*. - **LIBERO-Object** (10 tasks) — identical layout, a different object each time. - **LIBERO-Goal** (10 tasks) — identical objects and layout, a different goal each time. - **LIBERO-Long / LIBERO-10** (10 tasks) — long-horizon tasks that entangle all three factors, e.g. *turn on the stove and put the moka pot on it*. - **LIBERO-90** (90 tasks) — a broad pretraining pool drawn from kitchen, living-room and study scenes. LIBERO-90 and LIBERO-Long together are referred to as LIBERO-100. **What is actually in the release.** Exactly 50 human-teleoperated demonstrations per task — verified file by file, not assumed — giving **6,500 demonstrations and 1,007,618 transitions** across 130 HDF5 files totalling 93.5 GiB. At the 20 Hz control rate that is about 14 hours of interaction. Tasks are defined in PDDL-style BDDL files, so the suite is designed to be extended rather than frozen. **Observations and actions.** Each timestep carries an `agentview_rgb` and an `eye_in_hand_rgb` frame at 128×128, plus `joint_states` (7), `gripper_states` (2), `ee_pos` (3), `ee_ori` (3) and `ee_states` (6). Actions are 7-D OSC_POSE deltas. Full MuJoCo simulator state (`states`, 92-D) and the per-demo `init_state` are stored alongside, so any episode can be replayed deterministically — which is what every downstream re-render depends on. **Which copy to use.** This is worth knowing before downloading 93 GiB. The canonical HDF5 release is the one described above. Most vision-language-action work does **not** train on it directly, but on OpenVLA's regenerated `modified_libero_rlds`, which replays the demos in simulation and drops no-op-only and unsuccessful episodes — retaining **1,693 of 2,000 episodes (84.7%)** across Spatial, Object, Goal and Long. NVIDIA's LeRobot v3 conversion applies the same filtering and additionally covers LIBERO-90, keeping **5,614 of 6,500 episodes (86.4%)** and 842,714 of 1,007,618 frames. Physical Intelligence's LeRobot v2 copy is the four-suite filtered set (1,693 episodes / 273,465 frames). Numbers reported against "LIBERO" are usually against a filtered subset, not this one. **Licensing.** The authors' README licenses the codebase under MIT and the datasets under **CC BY 4.0**; the Hugging Face data mirror is labelled Apache-2.0. The CC BY 4.0 statement is the one that speaks specifically to the data and is used here.

simulation · LIBERO is a benchmark for **knowledge transfer in lifelong robot learning**, and its defining feature is experimental control rather than scale. Three of its suites hold everything fixed except one factor, so a policy's failure can be attributed to a specific kind of distribution shift instead of to "the scene changed". **The five suites.** - **LIBERO-Spatial** (10 tasks) — identical objects, different spatial relationships. The instruction is what disambiguates them: *pick up the black bowl next to the cookie box* vs *…on the ramekin*. - **LIBERO-Object** (10 tasks) — identical layout, a different object each time. - **LIBERO-Goal** (10 tasks) — identical objects and layout, a different goal each time. - **LIBERO-Long / LIBERO-10** (10 tasks) — long-horizon tasks that entangle all three factors, e.g. *turn on the stove and put the moka pot on it*. - **LIBERO-90** (90 tasks) — a broad pretraining pool drawn from kitchen, living-room and study scenes. LIBERO-90 and LIBERO-Long together are referred to as LIBERO-100. **What is actually in the release.** Exactly 50 human-teleoperated demonstrations per task — verified file by file, not assumed — giving **6,500 demonstrations and 1,007,618 transitions** across 130 HDF5 files totalling 93.5 GiB. At the 20 Hz control rate that is about 14 hours of interaction. Tasks are defined in PDDL-style BDDL files, so the suite is designed to be extended rather than frozen. **Observations and actions.** Each timestep carries an `agentview_rgb` and an `eye_in_hand_rgb` frame at 128×128, plus `joint_states` (7), `gripper_states` (2), `ee_pos` (3), `ee_ori` (3) and `ee_states` (6). Actions are 7-D OSC_POSE deltas. Full MuJoCo simulator state (`states`, 92-D) and the per-demo `init_state` are stored alongside, so any episode can be replayed deterministically — which is what every downstream re-render depends on. **Which copy to use.** This is worth knowing before downloading 93 GiB. The canonical HDF5 release is the one described above. Most vision-language-action work does **not** train on it directly, but on OpenVLA's regenerated `modified_libero_rlds`, which replays the demos in simulation and drops no-op-only and unsuccessful episodes — retaining **1,693 of 2,000 episodes (84.7%)** across Spatial, Object, Goal and Long. NVIDIA's LeRobot v3 conversion applies the same filtering and additionally covers LIBERO-90, keeping **5,614 of 6,500 episodes (86.4%)** and 842,714 of 1,007,618 frames. Physical Intelligence's LeRobot v2 copy is the four-suite filtered set (1,693 episodes / 273,465 frames). Numbers reported against "LIBERO" are usually against a filtered subset, not this one. **Licensing.** The authors' README licenses the codebase under MIT and the datasets under **CC BY 4.0**; the Hugging Face data mirror is labelled Apache-2.0. The CC BY 4.0 statement is the one that speaks specifically to the data and is used here.

Open · CC-BY-4.0 · hdf5 · lerobot