ACT trained with LeRobot 0.6.1 on 500 scripted demonstrations of the Isaac Lab task Mahjong-Tray1RedFaces-SO101-v1: an SO-101 arm picks up a single Red Dragon tile laid anywhere on a foam pad and places it in a tray. LeRobot's ACT defaults (ResNet-18 ImageNet backbone, chunks of 100, batch 8, learning rate 1e-5), 200,000 steps; the branches step- hold a checkpoint every 25,000 steps. Trained in simulation only; it has not been run on a real arm. A 50,000-step run of the same setup succeeded in 3 of 50 episodes; this run asks whether more training helps. The training frames show mahjong tile faces adapted from by Cangjie6, licensed under endorsement by the source author is implied. The…
Open weights
lerobot
pi0.5 trained on with LeRobot 0.6.1 to pick up, among seven tiles laid anywhere on a foam pad, the mahjong tile an instruction names, and place it in a tray, with an SO-101 arm in the Isaac Lab task Mahjong-Tray7TwosFaces-SO101-v1. Instructions read Pick up the {name} tile and place it in the tray. Trained in simulation only; it has not been run on a real arm. It goes on from the step-002500 checkpoint of which held over the named tile in closed loop. A trace of it found it deciding its descent by its joint state alone: held a few degrees off the hover pose its pushed demonstrations went down from, it went down once given that pose's state, and changed nothing with either camera blacked…
Open weights
gemma
lerobot
Fine-tuned from tc-huang/pi05mahjongv1tray07twosdetour (revision step-015000) with LeRobot 0.6.1 on the LeRobot dataset mahjongv1tray07twosdetour (run pi0520261014031144), with the pi05 preset of scripts/trainlerobot.sh and --steps=15000 --policy.optimizerlr=1e-5 --policy.schedulerwarmupsteps=250 --policy.schedulerdecaysteps=15000 --policy.schedulerdecaylr=2.5e-6 --savefreq=5000 --logfreq=200 --policy.freezevisionencoder=false --policy.gradientcheckpointing=false --wandb.enable=false: in float32, the vision encoder frozen unless those arguments unfreeze it. Its preprocessor does not crop the cameras. State and actions are normalized by the quantiles of all the dataset's frames, not by the…
Open weights
lerobot
