pp-bench

Simulated pick-and-place demonstrations of RoboStudio's Galaxea R1 Pro (r1pro_rs, a driving base) on ManiSkill3 / SAPIEN, from HQ-Bench's PickPlace-v1. One LeRobot v3.0 dataset per task, in a folder named after it.

• 20 Hz, joint-position actions (23) and state (28); three cameras, head and both wrists, 224x224, path-traced (rt-med).
• Each episode has its own instruction ("Put the tissue box in the box.").
• Per task 100 episodes with no clutter (T0) and 25 at each clutter level T1, T2, T3 and TH.
• A demonstration is kept when the scripted planner solves it with RoboStudio's collision gate at "ok", the success held for 10 steps, and no arm or torso joint moving over 0.5 rad in a step.

folder	task	episodes
PP02	Put the {object} in the {box}.	200
PP05	Put the {object} on the top shelf.	200
PP08	Take the {object} from the shelf and put it on the table.	200
MS24	Put the bottles on the {shelf} and the cups in the {box}.	191 (16 TH)
MS25	Tidy up: put the toys in the {box} and throw the trash in the {bin}.	200

eval_seeds.json lists 21 evaluation setups per task (10 at T0, 3 each at T1-T3, 2 at TH), seeds from the evaluation bank, every one solved cleanly by the planner: gym.make("PickPlace-v1", task=..., clutter=...), then reset(seed=...).

At TH the way to the target is closed: an obstructor stands in its approach lane and one at its side, and the robot sets them aside first. (On PP08 the target starts on a shelf, so its TH is clutter without obstructors.)

Correction, 2026-11-09. Until this date most episodes carried another episode's instruction: the export labelled them in numeric trajectory order, the converter numbers them alphabetically (PP02 197/200 wrong, PP05 195/200, PP08 167/175, MS24 149/175, MS25 147/175). Every episode was relabelled from its own recording, matched by its actions; frames, actions and states were right and are unchanged. Data downloaded before then should be fetched again.