A Pragmatic VLA Foundation Model

AI summary

Overview

Research area: Robotics — Vision-Language-Action (VLA) foundation models for real-world robotic manipulation.

Technical level: Advanced. The paper combines large-scale real-world data collection, a Mixture-of-Transformers model architecture, flow-matching action generation, and distributed-training systems engineering.

Scope: The paper describes LingBot-VLA, a robot manipulation foundation model pre-trained on about 20,000 hours of real-world dual-arm robot data from 9 robot configurations, evaluated on the GM-100 real-world benchmark and the RoboTwin 2.0 simulation benchmark, together with an optimized training codebase.

What This Paper Is About

VLA foundation models promise robots that can follow natural-language instructions across many tasks and hardware platforms, but the field lacks large-scale empirical evidence on how real-robot performance scales with pre-training data, and it lacks fast, efficient training codebases for such large-scale experiments. The authors build LingBot-VLA to answer the question "How do VLA models truly scale with massive real-world robot data?" and to check whether the resulting model generalizes across many robot platforms at acceptable cost. They pre-train on roughly 20,000 hours of real-world manipulation data and test the model on 100 tasks across multiple robotic platforms, while also measuring training throughput against existing open-source VLA codebases.

Key Contributions

1. A large real-world pre-training corpus. LingBot-VLA is pre-trained on approximately 20,000 hours of teleoperated real-world data collected from 9 popular dual-arm robot configurations, including AgiBot G1, AgileX, Galaxea R1Lite, Galaxea R1Pro, Realman Rs-02, Leju KUAVO 4 Pro, Qinglong, ARX Lift2, and Bimanual Franka.

2. A systematic scaling study of real-world robot data. By scaling pre-training data from 3,000 hours to 20,000 hours, the authors report consistent and substantial gains in downstream success rates, with no sign of saturation even at the 20,000-hour mark.

3. A large-scale multi-platform evaluation. Using the GM-100 benchmark of 100 carefully designed tasks, the authors evaluate 25 physical robots spanning 4 distinct commercial platforms, comparing LingBot-VLA against three state-of-the-art VLA baselines (π0.5, GR00T N1.6, and WALL-OSS) under identical training and testing conditions.

4. An optimized training codebase. The released codebase reaches a throughput of 261 samples per second on an 8-GPU training setup, a 1.5~2.8× speedup over existing VLA-oriented codebases depending on the VLM base model used.

The authors also release code, base model, and benchmark data openly.

Main Findings

• Real-world benchmark superiority. On the GM-100 real-world benchmark, LingBot-VLA with depth reaches an average success rate (SR) of 16.87% and progress score (PS) of 35.16%, compared with 13.00% / 27.32% for π0.5, 7.31% / 16.66% for GR00T N1.6, and 3.85% / 10.70% for WALL-OSS.

• The "without depth" variant also wins. LingBot-VLA w/o depth reaches 16.20% SR and 34.32% PS on average, outperforming WALL-OSS and GR00T N1.6 on all platforms in both metrics.

• Depth adds a measurable gain over π0.5. Incorporating depth-based spatial information gives LingBot-VLA w/ depth an average SR improvement of 4.28% and a PS increase of 7.76% over π0.5 across the evaluated embodiments.

• Per-platform results vary by robot. Table 1 reports Agibot G1 (Wheeled), AgileX (Wheeled), Galaxea R1Pro (Wheeled), and Leju KUAVO 4 Pro (Bipedal) separately. For example, LingBot-VLA w/ depth reaches 20.98% SR / 35.40% PS on Galaxea R1Pro and 18.93% SR / 40.36% PS on AgileX, while its Leju KUAVO 4 Pro SR (15.60%) is slightly below the Agibot G1 SR (11.98%) comparison pattern, and is lower than π0.5's 17.59% SR / 36.22% PS on Leju KUAVO 4 Pro only for the w/ depth variant.

• Pre-training on a matching platform helps. GR00T N1.6 performs average on Agibot G1, AgileX, and Leju KUAVO 4 Pro, but achieves SR and PS comparable to π0.5 on Galaxea R1Pro, which the authors attribute to the extensive inclusion of Galaxea R1Pro data during its pre-training.

• Simulation results on RoboTwin 2.0. Across 50 representative manipulation tasks: in clean scenes, average SR is 82.74% for π0.5, 86.50% for LingBot-VLA w/o depth, and 88.56% for w/ depth; in randomized scenes, 76.76%, 85.34%, and 86.68% respectively. The authors state these correspond to absolute gains over π0.5 of 3.76% and 8.58% for the w/o depth model, and 5.82% and 9.92% for the w/ depth model, in clean and randomized configurations.

• Favorable scaling with data volume. Progress rate and success rate both trend upward from 3,000 to 20,000 hours of pre-training data, measured on a subset of 25 representative tasks, with individual trends on Agibot G1, AgileX, and Galaxea R1Pro generally aligning with the aggregate.

• Strong data efficiency after pre-training. On eight representative GM-100 tasks on the Agibot G1 platform, LingBot-VLA with only 80 demonstrations per task outperforms π0.5 using the full 130-demonstration set in both progress rate and success rate, and the margin widens as post-training data grows.

• High test-set diversity. Roughly 50% of atomic actions in the test set are absent from the top 100 most frequent training actions.

• Training throughput. The codebase reaches 261 samples per second on an 8-GPU setup, outperforming StarVLA, Dexbotic, and OpenPI baselines in both the Qwen2.5-VL-3B-π and PaliGemma-3B-pt-224-π settings, and scales close to the theoretical linear limit at 8, 16, 32, 128, and 256 GPUs.

Methodology in Plain English

The authors first assembled a large teleoperated dataset from nine different dual-arm robot setups, then labeled it: human annotators split multi-viewpoint videos into clips around atomic actions and trimmed static start/end frames, and Qwen3-VL-235B-A22B was used to write precise task and sub-task instructions.

The model pairs a pre-trained vision-language model (Qwen2.5-VL) with a separately initialized "action expert," connected through a Mixture-of-Transformers design in which vision-language and action pathways run in separate transformer streams but share self-attention, so semantic knowledge from the VLM guides action generation at every layer while limiting cross-modal interference. At each timestep the model consumes three-view images, the task instruction, and the robot's proprioceptive state, and predicts a chunk of 50 future actions using flow matching, which models a smooth continuous path from noise to the ground-truth action. Blockwise causal attention keeps future action tokens from leaking into the observation representation.

To give the model spatial awareness, learnable queries tied to the three camera views are aligned with depth tokens from a depth model called LingBot-Depth through a distillation loss, injecting geometric information. On the engineering side, the team used Fully Sharded Data Parallel with dedicated shard groups for the action expert to cut communication overhead, mixed-precision training with float32 reductions for numerical stability and bfloat16 for storage and communication, plus FlexAttention and operator fusion to speed up the sparse attention computation.

For evaluation, they collected 150 raw trajectories per GM-100 task across three platforms, kept the top 130 by execution quality, standardized objects per GM-100 specifications, randomized object poses, and applied automated filtering plus manual human review. All baselines were fine-tuned from public checkpoints with the same post-training pipeline (batch size 256, 20 epochs), tested on the exact robot units used for data collection, in randomized order, with 15 trials per task-robot pair. Performance was scored by success rate within a 3-minute limit and by progress score over sequential subtask checkpoints.

Why This Matters

Impact on research. The paper provides rare empirical evidence that real-world VLA performance improves steadily with pre-training data volume up to 20,000 hours without saturating, which is a direct argument for continued large-scale real-world data curation rather than simulation-only scaling. It also introduces an evaluation protocol (100 tasks, multiple embodiments, 130 post-training episodes per task, 15 trials per task-robot pair, randomized object placement) intended as a standard for sound VLA benchmarking. The released code, base model, and benchmark data lower the barrier for reproducing and extending this work.

Real-world applications (as enabled by general manipulation skills):

• Tabletop dual-arm manipulation such as stacking bowls, which appears explicitly in the paper's task list.
• Cross-platform deployment, since the same model transfers to wheeled dual-arm robots and a bipedal humanoid.
• Fast task adaptation in industry settings, where the reported data efficiency (80 demonstrations per task beating a baseline trained on 130) reduces the demonstration burden per new task.
• Depth-aware spatial tasks that require precise geometric reasoning.

Industry relevance. The measured training throughput of 261 samples per second on 8 GPUs and the 1.5~2.8× speedup over existing VLA codebases translate directly into lower GPU-hour costs for companies training or fine-tuning robot policies, which the authors explicitly frame as making the model well-suited for real-world deployment.

Future Directions

• Extending beyond dual-arm tabletop setups. The authors state that future work will integrate single-arm and mobile robotic data to enable more diverse manipulation in unconstrained environments.
• Investigating whether scaling continues to help. The paper reports no saturation at 20,000 hours, leaving open how far the trend extends and whether it holds with additional embodiments.
• Understanding why depth helps and where it does not. The w/ depth variant is not uniformly better on every platform row in Table 1 (for example, Leju KUAVO 4 Pro SR is 15.60% w/ depth versus 17.59% w/o depth), suggesting follow-up study of when spatial distillation pays off.
• Broadening evaluation. The paper notes that real-world evaluation is bottlenecked by hardware parallelism and that prior work was confined to a few methods and tasks; expanding the benchmark to harder tasks and more platforms is presented as an open priority.

Target Audience

Robotics and embodied-AI researchers working on VLA foundation models and manipulation policies; machine learning systems engineers interested in distributed training efficiency for multimodal models; evaluation and benchmarking researchers who need multi-platform, large-scale real-world protocols; and industry teams deciding whether large-scale real-world robot data collection is worth the cost, since the paper quantifies both the scaling benefit and the training overhead.

Authors’ abstract

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.