FlexPhysics/sharpa-cylinder-rotation

Sharpa rotate dataset (format v1)

In-hand cylinder rotation trajectories collected in FlexPhysics with a frozen SharpaWave ProprioAdapt policy. Each NPZ is one episode.

Repository layout and Hub preview

This repository contains eight cylinder-scale subsets: scale_0.400, scale_0.429, scale_0.457, scale_0.486, scale_0.514, scale_0.543, scale_0.571, and scale_0.600. Each currently contains 128 episode NPZs (1,024 episodes in total). Folder names round the exact scale; use object_scale in each manifest for its precise value.

README.md
scale_*/
  metadata.json      # Run settings and NPZ schema
  manifest.jsonl     # One row per episode
  episodes/*.npz    # State/action trajectories
  videos/*.mp4      # Rendered replays
  videos/metadata.jsonl  # HF VideoFolder browsing index

The YAML configuration selects MP4 files and their VideoFolder metadata, with one Hub configuration per scale. After upload and Hub processing, the Dataset Viewer can show a playable video column alongside episode ID, scale, action count, validity, termination reason, seed, and grasp information. Select a scale_* configuration and the train split to browse its 128 videos. The train split is a storage designation, not a curated evaluation partition.

Each videos/metadata.jsonl is a derived browsing index. Its file_name is relative to that index's directory; npz_path and manifest_path are relative to the repository root. Original manifests remain the complete episode records, and metadata.json retains the run settings. The Viewer does not expand NPZ arrays or automatically join these original files. Episode IDs are local to a scale; use (scale_subset, episode_id) as a cross-subset identifier.

Upload the contents of this directory with this README at the Hub repository root, preserving all subdirectories. Upload MP4s and browsing indexes together. If episodes are added or recollected, regenerate the corresponding browsing index from the manifest. This is the native NPZ format with a video browsing index, not a LeRobot v3 export.

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

Generation

• Command seed: 0
• Checkpoint: learning/sharpa_rotate/assets/0.4-0.6-8.pth
• Grasp cache: learning/sharpa_rotate/assets/sharpa_grasp_linspace_0.4-0.6-8.npy
• URDF: examples/urdf/sharpa/wave_01/right_sharpa_wave/right_sharpa_wave.urdf
• Control dt: 0.05 s (20.0 FPS)
• Simulation dt: 0.004166666666666667 s, decimation 12
• Settle steps before t=0: 24
• Vector slots: 128

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

Randomization

Size buckets and grasps are sampled from the official grasp cache. Elastomer friction is sampled per episode from (0.2, 1.0) using the episode seed (not in the grasp cache). Object mass, object friction, and metal friction stay at evaluator constants (OBJECT_MASS=0.05, OBJECT_FRICTION=0.5, metal 0.1).

manifest.jsonl is JSON Lines: one compact JSON object per episode so rows can be appended and flushed without rewriting the whole file. Pretty-printed JSON for the run lives in metadata.json.

Height failure uses |z - z_reset| > 0.02 m. That is not labeled success, slip, or drop.

reward from the env is always 0 and priv_info is zeros; do not use them as labels. Object state and the recorded physical parameters are the privileged targets.

Videos

Videos are generated from recorded episodes at the control rate (20 FPS). Each MP4 can be reconstructed by replaying the corresponding NPZ. Each video frame corresponds to the same instant as state[t] in the NPZ. Physics substeps within a control step are not shown. Playback time matches simulation time:video_time = frame_index * control_dt.

• video frame t ↔ NPZ state[t] / time[t]
• T actions → T+1 video frames
• last frame is held for one extra 1/fps in the player

Assets and license

Checkpoint, grasp cache, and policy code are vendored from the SharpaWave / RMA in-hand rotation stack. URDF assets live under assets/.