shubham4413/so101_sim

SO-101 Gazebo Nut-Stacking Demonstrations

This repository contains LeRobot v3 imitation-learning demonstrations for an SO-101 arm in Gazebo. A ground-truth inverse-kinematics expert performs:

Pick up the red nut and stack it on the other red part.

The demonstrations cover approach, grasp, lift, transfer, placement, release, and retreat under randomized source and destination poses.

Available configurations

The repository contains two self-contained LeRobot dataset layouts:

Configuration	Repository path	Episodes	Frames	FPS	Approx. duration
verified_100	repository root	100	190,176	30	105.7 min
improved_50	50_train_improved/	50	91,629	30	50.9 min

Both configurations contain one task and use the so101_gazebo robot type. They are separate dataset layouts; the 50-episode directory should not be interpreted as a declared evaluation split of the root dataset.

Observations and actions

Every frame contains:

• observation.images.left: 640×480 RGB left-view video, AV1, 30 FPS
• observation.images.fpv: 640×480 RGB wrist/FPV video, AV1, 30 FPS
• observation.state: six measured joint positions
• action: six trajectory-controller joint-position references
• observation.environment_state: eight privileged simulator values
• timestamps and global, episode, frame, and task indices

The six state/action dimensions are shoulder pan, shoulder lift, elbow flex, wrist flex, wrist roll, and gripper position. The eight environment-state values are XYZ and yaw for the randomized source and destination parts.

observation.environment_state is intended for analysis and validation. A policy intended for real-robot transfer should not depend on these privileged simulator-only values.

Data generation

Episodes were generated in Gazebo using a numerical inverse-kinematics expert. The source nut and destination part were randomized within reachable workspace ranges. Candidate layouts were checked for separation and IK feasibility before execution. Failed controller rollouts were discarded, and episodes were saved only after the final object pose satisfied the stack validation criterion.

Collection used two synchronized RGB cameras, measured joint feedback, and position-controller references. The videos and trajectory tables preserve the original 30 Hz timeline.

Loading

Load the root 100-episode layout directly with LeRobot:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("shubham4413/so101_sim")
print(dataset.num_episodes, dataset.num_frames)

The nested 50-episode layout can be used after downloading the repository and passing 50_train_improved/ as the local dataset root.

Intended uses

• imitation-learning and action-chunking experiments;
• offline robot-learning analysis;
• camera/state synchronization and data-pipeline testing;
• controlled simulation studies and reproducibility work.

Limitations

• This is synthetic Gazebo data, not real-robot data.
• Simulator visuals, contacts, actuator dynamics, latency, and calibration do not fully reproduce physical hardware.
• The data covers one robot, one workspace, one object family, and one task.
• The repository does not provide a benchmark evaluation protocol or statistically meaningful policy success rate.
• The declared train ranges describe storage splits; users should create episode-disjoint validation and test sets for model evaluation.
• Models trained on this data may fail under new viewpoints, lighting, geometries, dynamics, or hardware calibration.

Safety

Simulation performance is not evidence of safe physical deployment. Before using a trained policy on hardware, validate observation keys and calibration, enforce conservative joint and velocity limits, reject stale observations, use command clamping and collision-aware supervision, retain an emergency stop, and begin with non-actuating dry runs.

Privacy

The image observations are generated in simulation and do not intentionally contain people or personal information.

Related model

See shubham4413/so100_act_sim for a related ACT simulation checkpoint and rollout. Its recorded training configuration references a later 100-episode improved dataset with 183,023 frames, so it should not be assumed to be byte-identical to either layout in this repository.