LingBot-VLA has focused on Pragmatic:
- Large-scale Pre-training Data: 20,000 hours of real-world data from 9 popular dual-arm robot configurations.
- Strong Performance: Achieve clear superiority over competitors on simulation and real-world benchmarks.
- Training Efficiency: Represent a 1.5 βΌ 2.8Γ (depending on the relied VLM base model) speedup over existing VLA-oriented codebases.
- [2026-01-27] LingBot-VLA Technical Report is available on Arxiv.
- [2026-01-27] Weights and code released!
Requirements
- Python 3.12.3
- Pytorch 2.8.0
- CUDA 12.8
# Install Lerobot
pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/huggingface/lerobot.git
cd lerobot
git checkout 0cf864870cf29f4738d3ade893e6fd13fbd7cdb5
pip install -e .
# Install flash attention
pip install /path/to/flash_attn-2.8.3+cu12torch2.8cxx11abiTRUE-cp312-cp312-linux_x86_64.whl
# Clone the repository
git clone https://github.com/robbyant/lingbot-vla.git
cd lingbot-vla/
git submodule update --remote --recursive
pip install -e .
pip install -r requirements.txt
# Install LingBot-Depth dependency
cd ./lingbotvla/models/vla/vision_models/lingbot-depth/
pip install -e . --no-deps
cd ../MoGe
pip install -e .We release LingBot-VLA pre-trained weights in two configurations: depth-free version and a depth-distillated version.
- Pretrained Checkpoints for Post-Training with and without depth
| Model Name | Huggingface | ModelScope | Description |
|---|---|---|---|
| LingBot-VLA-4B Β | π€ lingbot-vla-4b | π€ lingbot-vla-4b | LingBot-VLA w/o Depth |
| LingBot-VLA-4B-Depth | π€ lingbot-vla-4b-depth | π€ lingbot-vla-4b-depth | LingBot-VLA w/ Depth |
To train LingBot with our codebase, weights from Qwen2.5-VL-3B-Instruct, MoGe-2-vitb-normal, and LingBot-Depth also need to be prepared.
- Run Command:
python3 scripts/download_hf_model.py --repo_id robbyant/lingbot-vla-4b --local_dir lingbot-vla-4b -
Data Preparation: Please follow RoboTwin2.0 Preparation
-
Training Configuration: We provide the mixed post-training configuration in five RoboTwin 2.0 tasks ("open_microwave" "click_bell" "stack_blocks_three" "place_shoe" "put_object_cabinet").
Click to expand full YAML configuration
model:
model_path: "path/to/lingbot_vla_checkpoint" # Path to pre-trained VLA foundation model (w/o or w depth)
tokenizer_path: "path/to/Qwen2.5-VL-3B-Instruct"
post_training: true # Enable post-training/fine-tuning mode
adanorm_time: true
old_adanorm: true
data:
datasets_type: vla
data_name: robotwin_5_new
train_path: "path/to/lerobot_merged_data" # merged data from 5 robotwin2.0 tasks
num_workers: 8
norm_type: bounds_99_woclip
norm_stats_file: assets/norm_stats/robotwin_50.json # file of normalization statistics
train:
output_dir: "path/to/output"
loss_type: L1_fm # we apply L1 flow-matching loss in robotwin2.0 finetuning
data_parallel_mode: fsdp2 # Use Fully Sharded Data Parallel (PyTorch FSDP2)
enable_full_shard: false # Don't apply reshare after forward in FSDP2
module_fsdp_enable: true
use_compile: true # Acceleration via torch.compile
use_wandb: false
rmpad: false
rmpad_with_pos_ids: false
ulysses_parallel_size: 1
freeze_vision_encoder: false # ViT need to be optimized
tokenizer_max_length: 24 # token numbers of task prompt
action_dim: 14 # Target robot action space dimension
max_action_dim: 75 # action dim in LingBot-VLA
max_state_dim: 75 # state dim in LingBot-VLA
lr: 1.0e-4
lr_decay_style: constant
num_train_epochs: 69 # finetuning 20k step
micro_batch_size: 32
global_batch_size: 256
max_steps: 220000
ckpt_manager: dcp
save_steps: 220000
save_epochs: 69
enable_fp32: true
enable_resume: true # resume training automatically
# ===========================================================================
# Depth Injection Parameters
# (Required only for LingBot-VLA with Depth. Ignore if not using depth)
# ===========================================================================
align_params:
mode: 'query' # Query-based distillation
num_task_tokens: 8 # Number of learnable task-specific tokens
use_image_tokens: True
use_task_tokens: False
use_text_tokens: False
use_contrastive: True
contrastive_loss_weight: 0.3
depth_loss_weight: 0.002
llm: # VLM Projection Settings
dim_out: 2048
image_token_size: 8
image_input_size: 224
depth:
model_type: MoRGBD
moge_path: /"path/to/moGe-2-vitb-normal"
morgbd_path: "path/to/LingBot-Depth"
num_layers: 1
num_heads: 4
dim_head: 32
ff_mult: 1
num_backbone_tokens: 256
token_size: 16
dim_out: 1024
input_size: 224
visual_steps: 10000
visual_dir: "path/to/output/images" # visualization path of depth distillation- Run Command:
# without detph
bash train.sh tasks/vla/train_lingbotvla.py ./configs/vla/robotwin_load20000h.yaml --model.model_path /path/to/LingBot-VLA --data.train_path path/to/mixed_robotwin_5tasks --train.output_dir /path/to/lingbot_robotwin5tasks/ --model.tokenizer_path /path/to/Qwen2.5-VL-3B-Instruct --train.micro_batch_size ${your_batch_size} --train.global_batch_size ${your_batch_size * your_gpu_num}
# with depth
bash train.sh tasks/vla/train_lingbotvla.py ./configs/vla/robotwin_load20000h_depth.yaml --model.model_path /path/to/LingBot-VLA-Depth --data.train_path /path/to/mixed_robotwin_5tasks --train.output_dir /path/to/lingbot_depth_robotwin5tasks --model.tokenizer_path /path/to/Qwen2.5-VL-3B-Instruct --model.moge_path /path/to/moge2-vitb-normal.pt --model.morgbd_path /path/to/LingBot-Depth-Pretrained --train.micro_batch_size ${your_batch_size} --train.global_batch_size ${your_batch_size * your_gpu_num}- Evaluation
# robotwin2.0
export QWEN25_PATH=path_to_Qwen2.5-VL-3B-Instruct
python -m deploy.lingbot_robotwin_policy \
--model_path path_to_your_model \
--use_length 50 \
--port port- Customized Post-training: To construct post-training in specified downstream tasks, we have provided an example and please refer to Custom for details.
Our LingBot-VLA achieves state-of-the-art results on real-world and simulation benchmarks:
- GM-100 across 3 robot platforms
| Platform | WALL-OSS | GR00T N1.6 | Ο0.5 | Ours w/o depth | Ours w/ depth | |||||
|---|---|---|---|---|---|---|---|---|---|---|
| SR | PS | SR | PS | SR | PS | SR | PS | SR | PS | |
| Agibot G1 | 2.99% | 8.75% | 5.23% | 12.63% | 7.77% | 21.98% | 12.82% | 30.04% | 11.98% | 30.47% |
| AgileX | 2.26% | 8.16% | 3.26% | 10.52% | 17.20% | 34.82% | 15.50% | 36.31% | 18.93% | 40.36% |
| Galaxea R1Pro | 6.89% | 14.13% | 14.29% | 24.83% | 14.10% | 26.14% | 18.89% | 34.71% | 20.98% | 35.40% |
| Average | 4.05% | 10.35% | 7.59% | 15.99% | 13.02% | 27.65% | 15.74% | 33.69% | 17.30% | 35.41% |
- RoboTwin 2.0 (Clean and Randomized)
| Simulation Tasks | Ο0.5 | Ours w/o depth | Ours w/ depth | |||
|---|---|---|---|---|---|---|
| Clean | Rand. | Clean | Rand. | Clean | Rand. | |
| Average SR | 82.74% | 76.76% | 86.50% | 85.34% | 88.56% | 86.68% |
If you find our work useful in your research, feel free to give us a cite.
@article{wu2026pragmatic,
title={A Pragmatic VLA Foundation Model},
author={Wei Wu and Fan Lu and Yunnan Wang and Shuai Yang and Shi Liu and Fangjing Wang and Shuailei Ma and He Sun and Yong Wang and Zhenqi Qiu and Houlong Xiong and Ziyu Wang and Shuai Zhou and Yiyu Ren and Kejia Zhang and Hui Yu and Jingmei Zhao and Qian Zhu and Ran Cheng and Yong-Lu Li and Yongtao Huang and Xing Zhu and Yujun Shen and Kecheng Zheng},
journal={arXiv preprint arXiv:2601.18692v1},
year={2026}
}This project is licensed under the Apache-2.0 License.
We would like to express our sincere gratitude to the developers of VeOmni and LeRobot. This project benefits significantly from their outstanding work and contributions to the open-source community.