Jiahao-Wang/lingbot-va-robotwin-8task-step10000

LingBot-VA RoboTwin 8-Task Post-Trained Checkpoint

This repository contains a complete LingBot-VA inference bundle for RoboTwin evaluation.

Checkpoint

• —Training run: RoboTwin 8-task post-training
• —Checkpoint step: 10000
• —Base checkpoint source: LingBot-VA base checkpoint
• —Post-training update: transformer weights only
• —Bundle layout: full inference bundle with tokenizer/, text_encoder/, vae/, and transformer/

Training Tasks

• —click_alarmclock
• —stack_blocks_two
• —handover_block
• —place_object_basket
• —pick_dual_bottles
• —shake_bottle
• —turn_switch
• —beat_block_hammer

Usage

Download the model:

huggingface-cli download Jiahao-Wang/lingbot-va-robotwin-8task-step10000 \
  --repo-type model \
  --local-dir /path/to/lingbot-va-robotwin-8task-step10000

Start the LingBot-VA server:

python -m torch.distributed.run --nproc_per_node 1 \
  wan_va/wan_va_server.py \
  --config-name robotwin \
  --pretrained-model-path /path/to/lingbot-va-robotwin-8task-step10000 \
  --port 29056

Run RoboTwin rollout evaluation on an x86_64 NVIDIA/Vulkan-capable machine with RoboTwin 2.0 and SAPIEN installed.

Notes

The DeltaAI GH200/aarch64 environment used for training can load and serve this checkpoint, but RoboTwin 2.0 rollout evaluation is blocked there by the lack of an upstream sapien==3.0.0b1 linux-aarch64 wheel.