Source-linked AI summary
A Pragmatic VLA Foundation Model
Wei Wu, Fan Lu, Yunnan Wang, Shuai Yang, Shi Liu, Fangjing Wang, Qian Zhu, He Sun, Yong Wang, Shuailei Ma, Yiyu Ren, Kejia Zhang, Hui Yu, Jingmei Zhao, Shuai Zhou, Zhenqi Qiu, Houlong Xiong, Ziyu Wang, Zechen Wang, Ran Cheng, Yong-Lu Li, Yongtao Huang, Xing Zhu, Yujun Shen, Kecheng Zheng
TL;DR
VLA models need to generalize across tasks and robotic platforms while controlling adaptation costs, but real-world scaling evidence is limited. LingBot-VLA combines about 20,000 hours of data from 9 dual-arm platforms with an integrated architecture and optimized training codebase; it reports clear superiority across its benchmark, 261 samples per second on 8 GPUs, and open releases for further evaluation.
Problem
Real-world VLA scaling across massive datasets remains insufficiently studied, while efficient codebases for such evaluations are lacking.
Method
LingBot-VLA combines large-scale real-world pre-training, a VLM–action-expert architecture, conditional flow matching, and efficiency-oriented training optimization.
Results
LingBot-VLA achieves clear superiority across 100-task evaluations, with 4.28% average SR and 7.76% average PS improvements over π0.5 across three embodiments.
Takeaways & Limitations
The model, code, and benchmark data are released to support real-world robotic applications, challenging tasks, and sound evaluation standards.
Abstract
from arXiv · showhide
Offering great potential in robotic manipulation, a capable Vision-Language-Action (VLA) foundation model is expected to faithfully generalize across tasks and platforms while ensuring cost efficiency (e.g., data and GPU hours required for adaptation). To this end, we develop LingBot-VLA with around 20,000 hours of real-world data from 9 popular dual-arm robot configurations. Through a systematic assessment on 4 robotic platforms, each completing 100 tasks with 130 post-training episodes per task, our model achieves clear superiority over competitors, showcasing its strong performance and broad generalizability. We have also built an efficient codebase, which delivers a throughput of 261 samples per second with an 8-GPU training setup, representing a 1.5~2.8$\times$ (depending on the relied VLM base model) speedup over existing VLA-oriented codebases. The above features ensure that our model is well-suited for real-world deployment. To advance the field of robot learning, we provide open access to the code, base model, and benchmark data, with a focus on enabling more challenging tasks and promoting sound evaluation standards.
1 Introduction
LingBot-VLA addresses limited evidence on how VLA performance scales with massive real-world datasets and presents a model, evaluation framework, and efficient codebase for deployment-oriented robot learning.
- Real-world VLA scaling remains underexplored, and efficient codebases for evaluating massive datasets are lacking.
- 20,000 hours of pre-training data improve downstream success rates consistently and substantially from 3,000 hours, without saturation at 20,000 hours.
- 130 episodes per task across 4 robotic platforms support systematic assessment emphasizing task diversity and multi-platform consistency.
- LingBot-VLA scales dual-arm real-world data for pre-training, transfers to downstream tasks, and is assessed across robotic embodiments.
- 261 samples per second on an 8-GPU cluster combines state-of-the-art performance, generalization, and computational efficiency.
2 Related Work
Related work positions VLA foundation models around larger and more diverse pre-training, spatial representations, evaluation methodology, and specialized training infrastructure.
- Foundation VLA: VLA foundation models use pre-trained vision-language backbones with diffusion-based action heads to support multi-task execution and multi-embodiment adaptability.
- Spatial VLA: Spatial VLA research adds spatial representations and depth information to address geometric reasoning and depth perception in complex manipulation.
- Evaluation Methodology: Robot-policy evaluation spans simulation and real-world embodiment benchmarks, trading rapid, low-cost testing against physical-world complexity.
- Training Infrastructure: Open-source training codebases target different priorities, including flexible frameworks, VLA–VLM co-training, and specialized infrastructure.
3 Pre-training Dataset
LingBot-VLA’s pre-training dataset combines large-scale teleoperated data from nine dual-arm embodiments with annotated action-centric video segments and instructions.
- Data Collection: Nine popular dual-arm robot embodiments provide the foundation for the large-scale teleoperated pre-training dataset.
- Embodiments: Leju KUAVO 4 Pro, Qinglong, and ARX Lift2 extend the dataset across bipedal humanoid and dual-arm configurations with varied camera setups.
- Annotation: Human annotators decompose multi-view robot videos into atomic-action clips and remove static start and end frames to reduce redundancy.
4 Model Training
LingBot-VLA combines a pre-trained VLM with an action expert, spatial distillation, flow-matching action generation, and efficiency-oriented training infrastructure.
- Architecture: A Mixture-of-Transformers architecture couples separate vision-language and action pathways through shared self-attention for unified sequence modeling.
- Architecture: The joint sequence concatenates multi-view images, task instructions, robot state, and an action chunk.
- Action Generation: The action expert models p(A_t|O_t) with conditional flow matching, interpolating ground-truth actions and Gaussian noise across a flow timestep.
- Spatial Distillation: Vision distillation aligns VLM queries with LingBot-Depth tokens through projected cross-attention, infusing geometric information for precise manipulation.
- Training Efficiency Optimization: Distributed sharding, FlexAttention, and operator fusion optimize memory use, sparse multimodal computation, kernel overhead, and bandwidth utilization.
5 Experiments
The experiments evaluate LingBot-VLA across multiple robot platforms, tasks, datasets, and metrics using controlled real-world and simulation protocols. They emphasize multi-embodiment generalization, data quality, and both complete-task success and partial progress.
- Evaluation Setup: 22,500 controlled trials compare LingBot-VLA with three state-of-the-art baselines across 100 diverse manipulation tasks and four robotic platforms.The evaluation uses 25 physical robots and identical training and testing conditions.
- Evaluation Setup: Four dual-arm platforms use wrist-mounted and head-mounted cameras, with tabletop tasks and fixed chassis and waist configurations.The platforms are AgileX, Agibot G1, Galaxea R1Pro, and Leju KUAVO 4 Pro.
- Data Collection: 150 raw trajectories per task are collected by teleoperation, with the top 130 retained using completion, smoothness, and protocol-adherence criteria.The standardized collection protocol targets data quality and environmental diversity.
- Data Diversity: Approximately 50% of test-set atomic actions are absent from the 100 most frequent training actions, creating a diverse generalization assessment.The action-distribution discrepancy is used to characterize test-set diversity.
- Metrics: Success Rate measures complete task execution within three minutes, while Progress Score measures partial completion through sequential subtask checkpoints.Trials terminate after three consecutive subtask failures or safety-critical events, and metrics are reported across tasks and platforms.
5.2 Comparison on Real-world Benchmark
On the real-world GM-100 benchmark, LingBot-VLA is compared with strong baselines across multiple embodiments using success-rate and progress-score metrics. The depth-enabled variant improves over π0.5, while the model without depth outperforms WALL-OSS and GR00T N1.6.
- Real-world Results: 4.28% average SR improvement and 7.76% average PS increase over π0.5 are achieved by LingBot-VLA w/ depth across three embodiments.The comparison is reported across the three embodiments in Table 1.
- Platform Variation: GR00T N1.6 reaches SR and PS comparable to π0.5 on Galaxea R1Pro, while performing average on Agibot G1, AgileX, and Leju KUAVO 4 Pro.The paper attributes the Galaxea result to extensive Galaxea R1Pro data in GR00T N1.6 pre-training.
5.3 Comparison on Simulation Benchmark
Simulation experiments on 50 RoboTwin 2.0 tasks compare LingBot-VLA with π0.5 under clean and highly randomized scenes. LingBot-VLA improves success rates, with depth providing additional gains through learnable query-based spatial alignment.
- Simulation Setup: 50 representative RoboTwin 2.0 manipulation tasks are evaluated using clean and highly randomized scenes.Training uses 2,500 clean-scene demonstrations and 25,000 randomized-scene demonstrations.
- Simulation Results: Over 3.76% and 8.58% absolute success-rate gains over π0.5 are achieved by LingBot-VLA w/o depth in clean and randomized environments.The two gains correspond to the clean and randomized configurations, respectively.
- Simulation Results: 5.82% and 9.92% absolute success-rate margins over π0.5 are achieved with depth in clean and randomized configurations, respectively.Learnable query-based alignment extracts spatial priors from the LingBot-Depth model.
- Training Comparison: The throughput comparison standardizes the π-like architecture and local batch size at 32 while contrasting StarVLA, Dexbotic, OpenPI, and the proposed codebase.The evaluation uses sample throughput as its primary metric and aligns the VLM implementations for direct comparison.
- Training Comparison: The proposed codebase achieves the fastest training speeds for both Qwen2.5-VL-3B-π and PaliGemma-3B-pt-224-π settings while closely following theoretical linear scaling.Throughput is compared across 8, 16, 32, 128, and 256 GPUs.
5.5 Ablation Studies
Scaling experiments show that increasing real-world pre-training data from 3,000 to 20,000 hours consistently improves success and progress rates without observed saturation. Data-efficient post-training further shows strong performance with fewer demonstrations.
- Scaling Experiments: Success rate and progress rate consistently increase as pre-training data grows from 3,000 to 20,000 hours.The scaling experiments use 25 representative benchmark tasks.
- Scaling Experiments: The observed scaling trends show no saturation at 20,000 hours and generally align across Agibot G1, AgileX, and Galaxea R1Pro.The paper describes the scaling law as robust rather than specific to a single platform.
- Data-efficient Analysis: With only 80 demonstrations per task, LingBot-VLA outperforms π0.5 trained with the full 130-demonstration set on eight Agibot G1 tasks.The comparison covers both Progress Rate and Success Rate, and the margin widens as post-training data increases.
6 Conclusion
LingBot-VLA is presented as achieving strong generalizability and training efficiency, with comprehensive evaluation across 100 tasks. The authors also release the code, model, and benchmark data, while identifying broader robotic data integration as future work.
- Figure 6 concerns LingBot-VLA post-training data efficiency.
- LingBot-VLA achieves superior generalizability and training efficiency through large-scale real-world data and an optimized codebase.
- The conclusion reports clear superiority over competitors across 100 tasks, highlighting strong performance and broad generalizability.
- The authors release the code, model, and benchmark data to foster open science.
A Experiment
The supplementary experiments compile detailed real-world GM-100 results across multiple robotic platforms and a simulation evaluation on RoboTwin 2.0. These tables support aggregated mean results reported in the main text.
- Tables S1 and S2 report GM-100 results for the AgileX wheeled humanoid robot.
- Tables S3 and S4 report GM-100 results for the AgibotG1 wheeled humanoid robot.
- Tables S5 and S6 report GM-100 results for the Galaxea R1Pro wheeled humanoid robot.
- Tables S7 and S8 report GM-100 results for the Leju KUAVO 4 Pro bipedal humanoid robot.
- Table S9 reports RoboTwin 2.0 simulation results under clean and randomized settings.