junn0103/turbovla-so101-multitask3

TurboVLA SO-101 multi-task (3 tasks, 306 episodes) — step 160,000

TurboVLA policy trained on a merged real SO-101 corpus: 306 episodes / 179,821 frames, single wrist camera, 6-D absolute joint-position actions, THREE distinct task strings:

1. "pick the pen and put it in basket" (216 eps)
2. "Place the toy blocks into the cardboard box." (40 eps)
3. "Pick up the cubes and place them in the bowl" (50 eps)

At inference, pass one of the three strings verbatim as the task.

• Policy: turbovla_so101 (lerobot 0.6.1 plugin), dinov2-base + BERT frozen backbones
• Recipe: batch 256, 160k steps total, LR 5e-5 (10k warmup, cosine decay), L1 loss, 1x H200
• This upload: checkpoint at step 160,000
• Curves: https://wandb.ai/junn0103-seoul-national-university/lerobot/runs/mqxmchkd