helen9975/lingbot-vla-v2-6b-molmoact2-yam

LingBot-VLA 2.0 (6B) fine-tuned on MolmoAct2-BimanualYAM

Full fine-tune of LingBot-VLA 2.0 (robbyant/lingbot-vla-v2-6b, native depth) on the MolmoAct2-BimanualYAM dataset: 720+ hours of bimanual tabletop manipulation on two YAM arms (32,246 episodes, 76M frames at 30 Hz).

Intermediate checkpoint. This is step 40,000 of a planned 100,000-step run. It has not been evaluated open-loop or on a robot yet.

Recipe

Follows the MolmoAct2-BimanualYAM fine-tuning recipe from the MolmoAct2 paper where it applies to LingBot-VLA:

Cameras	top, left wrist, right wrist (dataset keys top, left, right)
Instructions	per-episode annotated instructions (meta/tasks_annotated.parquet)
Actions	absolute joint pose, 14-D: per arm 6 joints + gripper
Action chunk	30 steps (1 s at 30 Hz)
Global batch / planned updates	128 / 100K
Normalization	1st/99th-percentile bounds (bounds_99_woclip), from a random 1% of frames
Optimizer	Muon, LR 5e-5 constant (MoE experts 1.41e-4), LingBot's post-training defaults
Auxiliary losses	LingBot-VLA 2.0's depth and future-video distillation, kept on
Hardware	2 nodes x 8 H100, ~1.6 s/step

The normalization and optimizer aren't specified in the paper's recipe, so these use MolmoAct2's training-code default and LingBot's defaults respectively.

Training loss (flow matching, mean over ~200 steps): 0.115 at the start, 0.032 at 10K, 0.031 at 20K, 0.029 at 30K, 0.029 at 40K.

Files

checkpoints/global_step_40000/hf_ckpt/      # model weights + tokenizer/processor config
lingbotvla_cli.yaml                         # resolved training config (read by the policy server)
configs/robot_configs/molmoact2_yam_abs.yaml  # maps the dataset's 14-D state/action and cameras
assets/norm_stats/molmoact2_yam_abs.json      # normalization statistics

This mirrors the layout of upstream LingBot-VLA releases. Paths to the training-only teacher models in the configs are placeholders (/path/to/...); inference doesn't load them.

Usage

Set up the LingBot-VLA 2.0 repo and its environment (bash tools/create_train_env.sh), then download this repo and copy the robot config and norm stats into the code checkout:

huggingface-cli download helen9975/lingbot-vla-v2-6b-molmoact2-yam --local-dir lingbot-yam
cp lingbot-yam/configs/robot_configs/molmoact2_yam_abs.yaml configs/robot_configs/
mkdir -p assets/norm_stats && cp lingbot-yam/assets/norm_stats/molmoact2_yam_abs.json assets/norm_stats/
export QWEN3VL_PATH=/path/to/Qwen3-VL-4B-Instruct

Policy server:

python -m deploy.lingbot_vla_v2_policy \
  --model_path lingbot-yam/checkpoints/global_step_40000/hf_ckpt \
  --use_compile --use_length 30 --port 8000

Open-loop evaluation against recorded episodes:

python scripts/open_loop_eval.py \
  --model_path lingbot-yam/checkpoints/global_step_40000/hf_ckpt \
  --robo_name molmoact2_yam_abs \
  --data_path /path/to/MolmoAct2-BimanualYAM-Dataset \
  --use_length 30

Decoding the dataset's AV1 videos needs FFmpeg libraries for torchcodec (e.g. conda install -c conda-forge ffmpeg=7).

Observations are the three camera images (top, left wrist, right wrist), the 14-D joint state in the dataset's order (left_joint_0..5, left_gripper, right_joint_0..5, right_gripper) and a language instruction. The model returns a 30-step chunk of absolute joint positions in the same order.

Citations

@article{lingbotvla2,
  title={From Foundation to Application: Improving VLA Models in Practice},
  author={Wei Wu and Fangjing Wang and Fan Lu and He Sun and Shi Liu and Yunnan Wang and Yibin Yan and Yong Wang and Shuailei Ma and Xinyang Wang and Yibin Liu and Shuai Yang and Tianxiang Zhou and Kejia Zhang and Lei Zhou and Cheng Su and Nan Xue and Bin Tan and Han Zhang and Youchao Zhang and Fei Liao and Xing Zhu and Yujun Shen and Kecheng Zheng},
  journal={arXiv preprint arXiv:2607.06403},
  year={2026}
}

@misc{fang2026molmoact2actionreasoningmodels,
  title={MolmoAct2: Action Reasoning Models for Real-world Deployment},
  author={Haoquan Fang and Jiafei Duan and Donovan Clay and Sam Wang and Shuo Liu and Weikai Huang and Xiang Fan and Wei-Chuan Tsai and Shirui Chen and Yi Ru Wang and Shanli Xing and Jaemin Cho and Jae Sung Park and Ainaz Eftekhar and Peter Sushko and Karen Farley and Angad Wadhwa and Cole Harrison and Winson Han and Ying-Chun Lee and Eli VanderBilt and Rose Hendrix and Suveen Ellawela and Lucas Ngoo and Joyce Chai and Zhongzheng Ren and Ali Farhadi and Dieter Fox and Ranjay Krishna},
  year={2026},
  eprint={2605.02881},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2605.02881},
}