so101_block_stack_sim

Scripted-oracle demonstrations of an SO-101 stacking one 30 mm block on another, in a ManiSkill3 cell with two cameras on the intrinsic calibrated from the physical arm. Collected by sim/scripts/collect.py in vla-test.

360 episodes in LeRobot v3.0 format at 10 Hz, 60 for each of the six colour orderings. observation.state and action are the five arm joints and the gripper, observation.images.wrist and observation.images.top 480x480 H.264.

Layouts are screened on state alone and only the ones the oracle stacks are rendered, so every episode is a success. An episode starts where a cycle leaves the arm, jaws open or shut.

meta/train_layouts.jsonl records the block spawns and the opening pose each episode started from, and meta/eval_layouts.jsonl 150 more, held out, to measure a policy on. Both are what the env's reset takes back, so an evaluation sets off from where a demonstration did.

meta/stats.json gives observation.state and action the arm's joint travel as their normalization range. pi0.5 spells the state out in its prompt as one of 256 bins across that range and does not clip, so the range has to hold every pose the arm can reach.

Revision v2 is the earlier collection, and is what drwahl/pi05_so101_block_stack_sim was trained on. main carries the wrist camera panned onto where the arm holds it, matching the physical rig.