junn0103/turbovla-so101-pen-basket-216ep

TurboVLA SO-101 — "pick the pen and put it in basket" (216 episodes)

TurboVLA policy (lerobot_policy_turbovla_so101) trained on a merged real SO-101 dataset: 216 episodes / 129,122 frames @ 30fps from six recording sessions (cyclic and varied pen placements), single wrist camera, 6-D absolute joint-position actions.

• —Policy: turbovla_so101 (0.2B total params, ~14M trainable)
• —Vision backbone: facebook/dinov2-base (frozen; ungated substitute for DINOv3)
• —Language backbone: google-bert/bert-base-uncased (frozen)
• —Task string: "pick the pen and put it in basket"
• —Training: 320k steps, batch 256 (paper's global batch), LR 5e-5, 10k warmup, cosine decay over 320k, L1 loss, 1× H200, 15h24m
• —Final train loss: 0.012 (this checkpoint = step 320,000)
• —Curves: wandb run
• —Dataset: junn0103/pen_picking_all_216

Intermediate checkpoints (80k/160k/240k) available locally for generalization comparison.

Usage

pip install lerobot_policy_turbovla_so101   # Python >= 3.12

lerobot-record \
  --robot.type=so101_follower ... \
  --policy.path=<this repo id>

Trained with lerobot-train --policy.type=turbovla_so101 on lerobot 0.6.1.