smartseed-sim-grasp-white-v2

Simulated demonstrations of the SmartSeed grasp task (white seeds in the white box (real source d2)), generated in a MuJoCo twin of the scene. They are not real recordings. The schema is the one of RafaelJaime/smartseed-grasp-d2-clean (codebase v3.0, 30 fps, robot_type so_follower, same keys, shapes and units), so the same training scripts work unchanged and real and simulated episodes can be mixed or fine-tuned across.

• Replaces v1, which was withdrawn: in v1 the moving jaw and the fingertip pads did not collide with box and tray walls, so the open jaw could pass through a seed-box wall and the episode taught an impossible motion (37 % of the kept v1 episodes have an arm geom more than 1 mm inside a wall). v2 fixes the collision model and the planner and keeps only collision-free episodes (see "Validity" below). The v1 repos were deleted from the Hub; they can be regenerated from git history (https://github.com/alvaromoli13/Robotluces, generator at commit 874e274, the last one before the wall fixes).
• Task: grasp the seed below the gripper
• Episodes: 100, frames: 29508; generator success rate 89 % (112 attempts)
• Generator: sim/make_dataset.py of https://github.com/alvaromoli13/Robotluces at commit a279c23 (the twin, the planner and the randomisation live there; sim/README.md, "Generating sim datasets")

Validity

Every episode is simulated under a wall monitor (sim/collision.py): after each physics step (3.3 ms) it reads the contacts between any arm geom (all links, both jaw hulls, both pads) and any box or tray wall. An episode is kept only if

1. the seed rose over 3 cm and is still in the jaws at the end,
2. no arm geom went deeper than 1 mm into a wall (contact dist), and no wall contact pushed with more than 2 N,
3. before it was recorded the planner checked the whole commanded motion against every wall (forward kinematics): the open jaws stay 5 mm clear of walls and neighbouring seeds from the descent to the end of the lift, the transit and the retract stay 2 cm clear, and a scene where no jaw yaw fits is redrawn.

Walls are solid for every jaw geom in this scene (the floors, mat and table are not: the jaws rest on them at the taught pose, see sim/README.md, "Collision model").

Generator statistics

attempts / kept	112 / 100
lifted but rejected for a wall collision	0
scenes redrawn before simulating: total	666
... no jaw yaw / path clears the walls	269
... seed out of the arm's reach	397

Features

key	shape	content
observation.state	6 float32	measured follower joints: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll in degrees, gripper 0-100
action	6 float32	commanded (leader) joint targets, same names and units
observation.images.top	480x640x3 video	the twin's top camera
observation.images.wrist	480x640x3 video	the twin's wrist camera

What an episode does

Like the real teleop episodes: rest at the taught act_start, open the gripper while moving above a seed, descend, close, lift, fold back to the rest pose with the seed in the jaws and hold there until frame 300 (real episodes are 278-300 frames). Motion is min-jerk interpolation with random timing, a slow hand wobble and a 1.5 mm aiming error. The grasp is plain contact physics (no weld, no attach). Only successful, collision-free episodes are kept (see Validity).

Randomisation (per episode)

what	bound
seeds per box / which one is grasped	2-4 of 4 / one of the colour's seeds
seed position	uniform on the box floor, 22 mm off the inner walls, at least 25 mm apart
box and tray pose	the two boxes and the 3x4 tray anywhere on the table the top camera sees (uniform position, any yaw, no overlap, 15 mm inside the view, centre within 36 cm of the arm, 10 cm clear of its base); placements the planner cannot reach, or where no jaw yaw clears the walls, are redrawn
light	intensity x0.7-1.3, direction tilted up to 25 deg
colours	+-8 % per channel on table, mat and cardboard (the twin has no textures)
top camera	+-5 mm, +-1 deg
wrist camera	+-2 mm on the gripper
start pose	act_start plus per-joint noise (pan +-4, wrist +-3 deg)

sim_generation.jsonl logs every attempt (kept, failed, unreachable) with its seed and target.

Limits

The twin is not the real scene: flat colours, no print on the mat, no lens distortion, estimated box and seed layout, untuned servo dynamics (sim/README.md, "What is weak"). A policy trained on this data alone will see a visual gap on the real arm.