KhalilGao/dynasoftgrasp-m1-softmouse-pick

DynaSoftGrasp M1 · soft-body mouse pick episodes

200 episodes (190 judged SUCCESS, see success_manifest.json) of a Franka Panda picking a FEM soft-body computer-mouse-sized mouse (ByteDance Seed3D-generated mesh, 10 kPa tissue-scale Young's modulus baked into the USD) on a flat table. Collected with Isaac Sim 4.5 / Isaac Lab 2.2.1 at 25 Hz, 480x360, 3 cameras, on top of the DynamicVLA collection pipeline.

Task prompt (stored per episode in language): "grasp a static soft mouse"

Episode schema (one HDF5 per episode)

key	shape	dtype	meaning
action	(T, 8)	f32	EE target [pos(3), quat wxyz(4)] in robot root frame + gripper opening fraction g
state	(T, 7)	f32	current EE pose [pos(3), quat wxyz(4)], robot root frame — same space as action[:7], no gripper
language	(1,)	bytes	task prompt (UTF-8)
joints	(T, 9)	f32	measured arm joints (7) + finger joints (2, m)
ee_pos / ee_quat	(T,3)/(T,4)	f32	measured EE pose (= columns of state)
object_pos / object_quat / object_vel	(T,3)/(T,4)/(T,3)	f32	soft-body COM pose/velocity (quat is identity: FEM body has no root orientation)
opst_cam_rgb / wrist_cam_rgb / side_cam_rgb	(T,360,480,3)	u8	RGB views (training uses opst + wrist)
*_cam_seg	(T,360,480,1)	u8	semantic seg (0 bg / 1 robot / 2 mouse)
sm_state	(T,)	i32	scripted state machine phase

Gripper channel: normalization contract

Data is stored RAW; normalize in your dataloader. action[:, 7] is an opening fraction g in [0, 1] (1 = open, 0 = close command), recorded as a smooth ramp (0.075/frame). Evaluation MUST de-normalize:

per-finger joint target [m] = CLOSE + g * (OPEN - CLOSE)
CLOSE = 0.0098   # anti-crush close point (19.6 mm fingertip gap)
OPEN  = 0.0400          # fully open (80 mm gap)

g=0 is deliberately NOT a 0 mm pinch: the target is soft (~25-32 mm grasp slice); the fingers stall at 21-25 mm on the flesh (visible squeeze).

Success criterion (v2, judged per episode, success_manifest.json)

An episode is SUCCESS iff all of:

1. lift sustained — object COM rise: tail mean (last T/10 frames) >= 0.12 m AND last frame >= 0.10 m;
2. gripper closed ON THE BODY while risen — close command (g=0) on >= 95% of risen frames AND measured finger gap in [20.5, 30] mm. The 20.5 mm floor rejects the empty pinch (fingers at the bare 19.6 mm close command with nothing between them while the kinematic pin lifts the body anyway — flesh stall clusters at 21.8-25 mm, empty at 19.6-19.9 mm; one such episode was caught and re-labeled FAIL in this manifest);
3. object follows the hand — mean |EE - COM| while risen <= 0.05 m;
4. frame-0 static — finite-diff |vel| <= 0.02 m/s (clean spawn).

Action-stream note (for training)

action[:7] is an ABSOLUTE EE pose target tracked by a weak PD controller (~13% of the gap per frame). Phase entries therefore STEP the command: every successful episode contains one ~38 mm z step at the GRASP->LIFT transition (and a larger one at descent entry). This is a property of the action space — upstream DynamicVLA's raw data carries 0.4 m steps — not data corruption; delta-ized training consumes it as a large-but-consistent delta.

Honesty notes

• The lift is realized via a kinematic-carry workaround (a thin slab of FEM nodes is pinned to the EE during LIFT), NOT contact-force closure. The squeeze before the carry IS real FEM contact (finger stall + visible deformation), but do not cite these episodes as force-closure grasps.
• object_vel (root_vel_w) carries a ~0.02-0.03 m/s bias for a settled FEM body; derive true motion from object_pos finite differences.
• Episode randomization: object position (~0.24 x 0.16 m), yaw +/-45deg, fixed 1.4x mouse scale. Seeds 1000..1199.

Built on DynamicVLA's collection pipeline (soft-body support via runtime monkeypatch, upstream untouched). Repo: https://github.com/Autumn1337/DynaSoftGrasp