File size: 3,704 Bytes
9089132 05fc9de 9089132 05fc9de 9089132 05fc9de | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 | ---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- lohrbench
- panda
- curobo
- manipulation
- long-horizon
---
# LongHorizon — LoHRbench CuRobo Demonstrations (joint-pose)
Success-only demonstration trajectories for **8 long-horizon manipulation tasks** from
[LoHRbench](https://github.com/) (ManiSkill3, Franka Panda), collected with the **CuRobo**
GPU motion planner.
Two representations of the **same** success trajectories are provided:
| Path | Format | Episodes | Notes |
|---|---|---|---|
| `data/`, `meta/` | **LeRobot v2** | 1,200 (capped 150/task) | training-ready (pi0 / LeRobot pipelines) |
| `hdf5_joint_pose/` | **ManiSkill HDF5** | 1,383 (all successes) | raw trajectories + filtered metadata JSON |
Both are joint-pose (8D `pd_joint_pos`) and success-only. The HDF5 set keeps every
successful episode (no 150/task cap), so a few tasks have more than the LeRobot set.
## Contents
- **Episodes:** 1,200 (150 per task × 8 tasks), **all successful**
- **Frames:** 3,311,118 @ **20 fps**
- **Robot:** Franka Panda (`panda_wristcam`)
- **Planner:** CuRobo (GPU MotionGen)
- **Action space:** 8D `pd_joint_pos` — 7 arm joint targets + 1 gripper command (absolute)
### Features (per frame)
| key | dtype | shape | source |
|---|---|---|---|
| `image` | image | (128, 128, 3) | base camera RGB |
| `wrist_image` | image | (128, 128, 3) | hand/wrist camera RGB |
| `state` | float32 | (9,) | robot qpos (7 arm + 2 finger) |
| `actions` | float32 | (8,) | `pd_joint_pos` target (7 arm + gripper) |
## Tasks and language instructions
| task | category | instruction |
|---|---|---|
| `stack_10_cube` | super_long_horizon | Stack 10 cubes together, start with red cube |
| `reverse_stack` | super_long_horizon | Reverse 10 stacked cubes in reverse order |
| `stack_active_exploration` | active_exploration | Open the drawer, pick up the cubes, and stack them together starting with the red cube |
| `pick_active_exploration` | active_exploration | Pick up the can, screwdriver, and cup out of the drawer |
| `fruit_placement` | tool_using | Place four strawberries into the target positions |
| `repackage` | tool_using | Put the cube into the bowl and stack the bowl on the plate |
| `cluttered_packing` | clutter | Put three cubes into the bowl |
| `stack_cube_clutter` | clutter | Stack 3 cubes together, start with red cube |
The per-frame language instruction is stored in the LeRobot `task` field.
### HDF5 success counts (`hdf5_joint_pose/`)
| task | success episodes |
|---|---|
| stack_10_cube | 150 |
| reverse_stack | 172 |
| stack_active_exploration | 300 |
| pick_active_exploration | 150 |
| fruit_placement | 150 |
| repackage | 150 |
| cluttered_packing | 161 |
| stack_cube_clutter | 150 |
| **total** | **1,383** |
Each `<name>_success.h5` holds only the successful `traj_*` groups; the sibling
`<name>_success.json` is the ManiSkill trajectory metadata trimmed to those episodes.
Episode success is taken from the reconciled per-episode label (`episodes[].success`),
not the unreliable per-step flag. HDF5 schema: `obs/agent/qpos` (T,9),
`obs/sensor_data/{base,hand}_camera/rgb` (T,128,128,3), `actions` (T-1,8).
## Notes
- Only **successful** episodes are included (per-episode success labels reconciled from the
ManiSkill collection metadata; the per-step success flag is not used).
- Actions are **absolute** joint targets. For pi0/pi0.5-style training, convert the 7 arm
joints to deltas relative to the current state and keep the gripper absolute.
- This release contains the **joint-pose (`pd_joint_pos`)** representation only (both LeRobot
and HDF5). A delta-EE (`pd_ee_delta_pose`) variant is not included here.
|