lingbot-va-robotwin-8task-v2-step10000

LingBot-VA RoboTwin 8-Task V2 Step 10000

This is a LingBot-VA inference bundle post-trained on eight RoboTwin tasks.

Checkpoint

• Base model: /work/hdd/bfxg/jwang87/lingbot-va-base
• Post-training run: robotwin_8task_v2_base
• Checkpoint step: 10000
• Training job: Slurm 2425436
• WandB run: robotwin_8task_v2_10k_bs8_bs1_accum2
• Effective batch size: 8 (4 GPUs * per-GPU batch size 1 * gradient accumulation 2)

Only the transformer/ weights were updated during post-training. The tokenizer/, text_encoder/, and vae/ directories are copied from the LingBot-VA base checkpoint so the downloaded directory can be passed directly to LingBot-VA inference.

Training Tasks

• click_bell
• stack_bowls_two
• handover_mic
• place_can_basket
• pick_diverse_bottles
• shake_bottle_horizontally
• open_microwave
• move_stapler_pad

press_stapler was intentionally held out for generalization testing.

Usage

huggingface-cli download Jiahao-Wang/lingbot-va-robotwin-8task-v2-step10000 \
  --repo-type model \
  --local-dir <target_dir>

Then use:

python -m torch.distributed.run --nproc_per_node 1 wan_va/wan_va_server.py \
  --config-name robotwin \
  --pretrained-model-path <target_dir> \
  --port 29056