ASKCOS: an open source software suite for synthesis planning
Paper • 2501.01835 • Published
Stage-annotated demonstration dataset for the PickPlaceCan task from robosuite, formatted in LeRobot v2.1 format for training SARM (Stage-Aware Reward Models).
| Property | Value |
|---|---|
| Task | Pick and place can |
| Robot | Panda (Franka Emika) |
| Environment | robosuite PickPlaceCan |
| Data source | robosuite multi-human (MH) demonstrations |
| Episodes | 300 |
| Total frames | 62,756 |
| FPS | 20 |
| Cameras | agentview (third-person), robot0_eye_in_hand (wrist) |
| Image resolution | 256 x 256 |
| Video codec | AV1 (libsvtav1) |
| LeRobot version | v2.1 |
| Feature | Type | Shape | Description |
|---|---|---|---|
state |
float32 | (9,) | EE position (3) + EE quaternion (4) + gripper qpos (2) |
actions |
float32 | (7,) | End-effector delta action (6 DOF + gripper) |
reward |
float32 | (1,) | Continuous progress label (see Reward Encoding below) |
agentview-images-rgb |
video | (256, 256, 3) | Third-person camera view |
robot0-eye-in-hand-images-rgb |
video | (256, 256, 3) | Wrist-mounted camera view |
Each stage corresponds to a task prompt. Frames are assigned to the stage they belong to — not edges/transitions.
| Stage | Prompt | Description |
|---|---|---|
| 0 | "robot approaching the can" | From episode start until gripper reaches the can |
| 1 | "robot grasping and lifting the can" | From contact through lifting off the table |
| 2 | "robot moving the can towards placement bin" | Transporting the grasped can toward the bin |
| 3 | "robot placing the can in bin" | Lowering/releasing the can into the bin until done |
The reward field encodes a continuous progress label for SARM training:
reward = stage + (frame_position_in_stage / total_frames_in_stage)
For example, if stage 0 has 60 frames:
SARM decodes this as:
floor(reward) = stage class (0, 1, 2, or 3)frac(reward) = intra-stage progress (0.0 to ~1.0)Reward range: [0.0, 4.0) — never reaches 4.0.
Use num_classes_sparse: 4 in SARM config.
Stages are detected via direct MuJoCo state reading from robosuite simulation:
sim.data.site_xpos[eef_site_id]sim.data.body_xpos[can_body_id]sim.data.contact geom-pair matchingThe HDF5 stores only integer stage labels per frame. The continuous reward is computed during LeRobot conversion.
from lerobot.common.datasets.rm_lerobot_dataset import FrameGapLeRobotDataset
dataset = FrameGapLeRobotDataset(
repo_id="poolvarine/SARM-robosuite-can-mh-stages",
episodes=list(range(300)),
n_obs_steps=4,
frame_gap=1,
max_rewind_steps=2,
image_names=["agentview-images-rgb"],
task_name="pick and place can",
)
sample = dataset[0]
# sample["state"] -> (n_obs_steps+max_rewind_steps+1, 9)
# sample["targets"] -> continuous progress labels for reward model training
# sample["agentview-images-rgb"] -> decoded video frames
general:
camera_names: ["agentview-images-rgb"]
state_dim: 9
model:
num_classes_sparse: 4
annotation_list: ["approaching", "grasping_lifting", "transporting", "placing"]
If you use this dataset, please cite the SARM paper and robosuite:
@inproceedings{robosuite2020,
title={robosuite: A Modular Simulation Framework and Benchmark for Robot Learning},
author={Zhu, Yuke and Wong, Josiah and Mandlekar, Ajay and others},
booktitle={arXiv preprint arXiv:2009.12293},
year={2020}
}