FlexPhysics/pick_and_place

Pick-and-place dataset (format v1)

Robot pick-and-place trajectories collected with a scripted task-space expert and joint-space PD control. Each NPZ is one episode. The current release contains one UR10 with Robotiq-85 dataset: 32 valid episodes, 32 successful episodes, and 32 rendered videos.

Repository layout and Hub preview

README.md
robots/<robot_id>/
  README.md
  metadata.json              # Collection configuration, schema, ordering, and units
  summary.json               # Episode and outcome counts
  manifest.jsonl             # One complete index row per episode
  files.jsonl                # Size and SHA-256 inventory
  assets/<robot>.urdf        # Robot description used by the dataset
  episodes/*.npz             # State and control trajectories
  videos/*.mp4               # Rendered pose replays
  videos/metadata.jsonl      # Hugging Face VideoFolder browsing index

The YAML configuration selects the videos and their VideoFolder metadata. After upload and Hub processing, the Dataset Viewer can show a playable video column with robot ID, episode ID, action count, validity, success, termination reason, outcome label, failure cause, seed, and pointers to the NPZ and manifest.

videos/metadata.jsonl is a derived browsing index. Its file_name is relative to the videos/ directory. npz_path and manifest_path are relative to the dataset root. manifest.jsonl remains the authoritative episode index, and metadata.json remains the authoritative schema and collection configuration. The Viewer does not expand NPZ arrays or automatically join manifest diagnostics.

The train split is a storage designation, not a curated training/evaluation partition. Preserve the directory structure when publishing. This is a native NPZ dataset with a VideoFolder browsing index, not a LeRobot export.

VideoFolder format: https://huggingface.co/docs/hub/datasets-video

Current dataset

• Robot: UR10 with Robotiq-85 two-finger gripper
• Root seed: 0
• Episodes: 32
• Valid: 32
• Successful: 32
• Vector environments: 32
• Control rate: 60 FPS (control_dt = 1/60 s)
• Physics rate: 240 Hz (physics_dt = 1/240 s)
• Object: 0.05 × 0.05 × 0.05 m, 0.08 kg
• Coordinate convention: metres, radians, seconds, +Z up, quaternions in wxyz order

Randomization

Sampling is deterministic per episode. The episode seed is root_seed + episode_id × 1,000,003; retries also include the attempt index. Slot scheduling and vector-environment completion order do not change sampled values.

The following values are randomized independently for every episode and recorded in manifest.jsonl under episode_spec:

value	distribution
Initial arm joints	Uniform jitter around the IK home solution: ±[0.08, 0.06, 0.06, 0.06, 0.05, 0.08] rad
Object X	Uniform in [0.52, 0.58] m
Object Y	Uniform in [0.12, 0.18] m
Goal X	Uniform in [0.37, 0.43] m
Goal Y	Uniform in [-0.18, -0.12] m

Object and goal Z are fixed at 0.025 m, half the cube height. Object yaw is fixed at 0 in this run. Object size, mass, friction, robot base pose, gripper parameters, PD gains, control timing, and task tolerances are not randomized. Samples violating joint limits or the minimum 0.15 m transfer distance are rejected and resampled; rejection records are retained in planning_rejections.

Trajectory fields

A trajectory with T actions stores T+1 measured states. Alignment is:

state[t] → action[t] → state[t+1]

The initial hold is included. State arrays use shape (T+1, ...); transition/control arrays use (T, ...). The same schema is machine-readable in metadata.json under npz.

field	shape	dtype	description
time	(T+1)	float32	Control time in seconds
joint_position	(T+1, 8)	float32	Measured joint angles in radians
joint_velocity	(T+1, 8)	float32	Measured joint velocities in rad/s
joint_target	(T+1, 8)	float32	Active absolute joint target
tcp_position	(T+1, 3)	float32	TCP position in metres
tcp_quaternion	(T+1, 4)	float32	TCP orientation, wxyz
object_position	(T+1, 3)	float32	Object position in metres
object_quaternion	(T+1, 4)	float32	Object orientation, wxyz
object_linear_velocity	(T+1, 3)	float32	Object linear velocity in m/s
object_angular_velocity	(T+1, 3)	float32	Object angular velocity in rad/s
goal_position	(T+1, 3)	float32	Episode placement target in metres
gripper_contact	(T+1, 2)	bool	Left/right configured finger contact
body_position	(T+1, 18, 3)	float32	Recorded body positions for replay
body_quaternion	(T+1, 18, 4)	float32	Recorded body orientations for replay, wxyz
action	(T, 8)	float32	Absolute joint-position command
control_mode	(T)	int32	0=open/move, 1=close, 2=hold
payload_mass	(T)	float32	Payload mass used by feedforward compensation, kg
pd_gains	(T, 8, 2)	float32	Applied (kp, kd) gains
feedforward_torque	(T, 8)	float32	Applied feedforward torque in N·m
expert_phase	(T)	int32	Index into metadata.json: expert_phases
terminated	(T)	bool	Terminal transition marker
truncated	(T)	bool	Timeout transition marker

Joint order is stored in metadata.json: joint_names. Body order is stored in metadata.json: body_names; consumers should not infer either order from the URDF.

Episode index and labels

Every manifest.jsonl row contains:

• Identity and storage: episode_id, seed, episode_path, num_actions.
• Primary labels: valid, success, termination_reason.
• Randomization: episode_spec.initial_joints, object pose, goal pose, episode seed, and attempt.
• Planning/validation: planning_rejections, validation_errors, and error.
• Video status: path, frame count, FPS, dimensions, replay source, and export error.
• Outcome label: outcome.label, outcome.category, outcome.final_phase, outcome.failed_checks, and outcome.failure_cause.
• Task diagnostics: maximum lift, final 3D/XY/height placement error, oriented-box bottom height, final object orientation error, final TCP position/orientation error, and final arm tracking error.
• Final checks: released, supported, slow, TCP clear, gripper contact, final linear speed, and final angular speed.

A successful episode must lift the object, release it, leave it supported near the goal, reach the stability thresholds, and clear the TCP. Final object orientation is recorded as a diagnostic but is not part of the success criterion.

Failure rows are retained. Possible termination or outcome labels distinguish planning failure, grasp failure, object drop/out-of-bounds, phase timeout, support timeout, invalid state, simulation error, and time limit. failure_cause contains a more specific derived cause when the recorded evidence supports one; otherwise it is null.

valid means the trajectory arrays satisfy the data contract. It does not mean the task succeeded.

Videos and indexes

Videos are rendered from recorded body poses at the control rate. Physics is not rerun during video generation.

• Video frame t corresponds to NPZ state[t] and time[t].
• T actions produce T+1 video frames.
• Playback time is frame_index × control_dt.
• Successful and failed valid trajectories are both eligible for video export.

videos/metadata.jsonl supports Hub browsing. files.jsonl inventories every published robot-dataset artifact except itself and records relative path, role, byte size, and SHA-256 checksum. summary.json provides aggregate episode, validity, success, termination, outcome, vector-environment, and execution-backend counts.

Assets

Each robot directory stores one robot description at assets/ur10_grip.urdf. Mesh files are not duplicated into the dataset. The recorded body poses and exported videos remain the authoritative visual trajectory data.