FuchangWang/sharpa_test1

Sharpa In-Hand Cylinder Rotation

Simulated in-hand cylinder rotation trajectories collected in FlexPhysics with a frozen SharpaWave ProprioAdapt policy and a right Sharpa Wave hand. Each episode contains joint states, actions, tactile signals, object motion, and a corresponding visualization video. Data format version: 1.

Dataset overview

This release contains 31 episodes and 31 videos, with 2,930 action transitions and 2,961 states. Episodes contain 1–200 actions (0.05–10 seconds of simulated transitions) at a control rate of 20 Hz.

All 31 manifest entries are marked valid. Of these, 19 end with height_fail and 12 with step_limit. Validity does not indicate task success. There are no predefined train, validation, or test splits. When creating splits, keep complete episodes together to avoid sharing adjacent frames across splits.

The num_episodes value of 200 in metadata.json is the requested collection count, not the number included in this release. Use manifest.jsonl as the index of included episodes.

Files and access

File or folder	Contents
README.md	Dataset description and field definitions
metadata.json	Collection settings, array schema, joint order, units, and source asset hashes
manifest.jsonl	One JSON object per episode with file references and episode metadata
episodes/	One NumPy NPZ archive per episode
videos/	One MP4 visualization per episode

Matching NPZ and MP4 files share a name, such as episode_000000.npz and episode_000000.mp4. The manifest's episode_path and video.path are relative to the dataset root. Each row also records the seed, number of actions, validity, termination reason, grasp selection, physical parameters, initial pose, and motion statistics.

Download the files from the repository's Files and versions tab. NPZ files contain named NumPy arrays with the schema below; MP4 files can be opened in a standard video player. The Hub Dataset Viewer is configured to display manifest.jsonl, with one row per episode under the episodes configuration. Its train split contains all included episodes and is a display/loading convention, not a predefined training partition. NPZ arrays remain separate downloadable files; the viewer displays episode metadata and file references rather than decoded trajectory arrays or embedded video players. The schema below describes the NPZ contents.

Collection settings

• Collection seed: 0
• Policy checkpoint: 0.4-0.6-8.pth
• Grasp cache: sharpa_grasp_linspace_0.4-0.6-8.npy
• Hand description: right_sharpa_wave.urdf
• Compute backend: CPU
• Control timestep: 0.05 s (20 Hz)
• Simulation timestep: 1/240 s, with 12 simulation steps per control step
• Settle steps before time zero: 24
• Parallel environment count: 1

Source asset names identify the collection inputs; the checkpoint, grasp cache, and URDF are not bundled in this dataset. Their SHA-256 hashes are recorded in metadata.json.

Fields

A trajectory with T actions stores T+1 states. Alignment is state[t] → action[t] → state[t+1]. Time starts at 0 after contact settle. The same schema is in metadata.json under npz (keys and shapes, no NPZ parse).

field	shape	dtype	notes
time	(T+1)	float32	seconds, world clock after contact settle
obs	(T+1, 192)	float32	policy observation before running-mean normalization
proprio_hist	(T+1, 30, 64)	float32	proprioceptive history before sa_mean_std
joint_position	(T+1, 22)	float32	joint angles in radians, policy joint order
joint_target	(T+1, 22)	float32	PD position targets in radians, policy joint order
tactile	(T+1, 5)	float32	thumb to pinky elastomer net-force norms, scaled then smoothed
object_position	(T+1, 3)	float32	object world position in meters
object_linear_velocity	(T+1, 3)	float32	object world linear velocity in m/s
object_angular_velocity	(T+1, 3)	float32	object world angular velocity in rad/s
object_quaternion	(T+1, 4)	float32	object body quaternion, wxyz
action	(T, 22)	float32	clipped policy action sent to env.step
terminated	(T)	bool	physical or invalid stop at this transition
truncated	(T)	bool	time or step-limit truncation at this transition

Actions are incremental joint-target commands: target += 0.041666666666666664 * action, then clipped to 0.9 of URDF limits. Tactile smoothing is smooth * latest + (1-smooth) * previous with CONTACT_SMOOTH=0.5.

Joint order

Policy order (not URDF tree order):

• right_index_MCP_FE
• right_middle_MCP_FE
• right_pinky_CMC
• right_ring_MCP_FE
• right_thumb_CMC_FE
• right_index_MCP_AA
• right_middle_MCP_AA
• right_pinky_MCP_FE
• right_ring_MCP_AA
• right_thumb_CMC_AA
• right_index_PIP
• right_middle_PIP
• right_pinky_MCP_AA
• right_ring_PIP
• right_thumb_MCP_FE
• right_index_DIP
• right_middle_DIP
• right_pinky_PIP
• right_ring_DIP
• right_thumb_MCP_AA
• right_pinky_DIP
• right_thumb_IP

Physical parameters and randomization

This release uses a fixed object scale of approximately 0.457143 (grasp-cache bucket 2). Initial grasps are sampled from that bucket. The cache contains other size buckets, but those sizes are not represented in this release.

Elastomer friction is sampled per episode from 0.2–1.0 using the episode seed. The realized value is recorded in elastomer_friction and physical_params.elastomer_friction in the manifest. Object mass is fixed at 0.05 kg, object friction at 0.5, and metal friction at 0.1.

Video alignment

Videos are 1280 × 720 MP4 visualizations at 20 FPS. Each frame corresponds to a recorded state after settling or after a control step. Intermediate physics substeps are not shown.

• Video frame t corresponds to NPZ state[t] and time[t].
• action[t] is applied between frames t and t+1.
• T actions correspond to T+1 states and T+1 video frames.
• Playback time is frame_index × 0.05 s; the final frame occupies one extra frame interval.

Intended use and limitations

The trajectories can support analysis of dexterous manipulation, tactile and proprioceptive observations, and object dynamics. They come from one frozen policy in simulation, with one hand, one object shape, and one size bucket. They do not establish performance on physical hardware or broader object sets.

height_fail means the object exceeded a vertical displacement band of 0.02 m relative to its reset height. step_limit means the configured action limit was reached. Neither is an explicit success, slip, or drop annotation. Short episodes and height failures are retained, so select episodes according to your use case.

The collection environment's reward is always zero and its privileged-observation placeholder is zero-filled; neither is included as a training label in the listed NPZ fields. Recorded object states and physical parameters provide targets for analyses that require privileged information.

Provenance and license

The collection checkpoint, grasp cache, and policy implementation originate from the SharpaWave / RMA in-hand rotation stack. The source hand assets carry third-party notices in NOTICE.txt.

A dataset license has not yet been specified. The licenses of the source code and assets should not be assumed to apply automatically to these generated trajectories.